WorksComputer Science
Language models represent more about emphasis than they can tell you
Abstract
Prosodic prominence can largely be predicted from the words of a sentence, but part of it depends on how the speaker reads it. We study how much of prominence language models recover from text, whether they can report it when asked, and what the speech signal adds. On read English speech labelled for word-level prominence, we compare frozen and fine-tuned text models, prompted large language models (LLMs) and speech models on the same words. All text models end within about one point of each other, at best 68.8% on the 3-way task, and largely make the same errors. With the prompts we tested, LLMs do no better than a word-majority baseline, even with 500 examples. Linear probes find more in their representations. Speech models reach 84.7%. We argue that text determines only part of prominence, and that the rest has to be heard. LLMs encode what the text determines, but prompting does not bring it out: they represent more about where the emphasis falls than they tell you.
The code, the results of every run and the commands that reproduce them are publicly available.1
Introduction
When a sentence is read aloud, some of its words stand out more than others. Following Talman et al. [1], we treat this prosodic prominence as the degree of emphasis a word receives from the speaker, not as contrastive or emphatic focus in the narrow sense. Prominence has no formal definition that holds across languages, and word stress and phrasal accent are distinct [2], so our definition is operational: prominence is what the labels measure. Much of this is predictable from the text: content words are more often prominent than function words, new information more often than given information, and the last content word of a phrase tends to carry the main accent [3, 4]. However, prominence also depends on what the speaker wants to convey, which led Bolinger to remark that accent is predictable only if you are a mind-reader [5]. Speakers agree on accent placement only some of the time [6].
Large language models (LLMs) are now used to choose which words to emphasise in speech synthesis [7], continuing a line of work that uses text representations to predict prominence in synthesis front-ends [8, 9]. It is reasonable to expect that models trained on very large amounts of text have learned a great deal about information structure, and therefore about prominence. Do their representations encode where the emphasis falls? And if they do, can they tell us?
In this paper we answer these questions with a comparison of text and speech models on the prominence annotations of the LibriTTS audiobooks released by Talman et al. [1]. Since the labels were computed automatically from read speech, text models, prompted LLMs and speech models can be evaluated
on the same labels and on the same words. Our contributions are the following:
LLMs represent more about prominence than they can report. A linear probe on the hidden states of Qwen3.8 27B reaches 66.7% on the 3-way task, against 62.9% on its embedding layer, which encodes each word without its context; the contextual representations thus carry information about prominence beyond the identity of the word. Prompted for the same labels, with the prompts we tested, the model reports less: 58.6% without examples and 62.0% with 500 examples in the prompt, no better than a word-majority baseline that gives each word the label it most often has in the training data.
One ceiling for all text models. Text models of very different size and design, probed or fine-tuned, converge on the same accuracy and largely on the same errors, and a classifier over random word vectors comes close to it, which suggests that, for the models and data we tested, the limit lies in the text rather than in the models.
What the speech signal adds. Speech models, which hear how the sentence was actually read, are 16.5 points more accurate than the best text model on the same words. The gain is largest on content words and on the last word of the utterance, and when two speakers read the same text, the speech models show the one they hear.
A ceiling for speech models. The labels were computed automatically from the recordings. Running the same procedure again on public word alignments reproduces the original label for 91.5% of the words (3-way), which shows how far a speech model could go on these labels; the best speech model is still 6.9 points below this ceiling.
Related work
Prominence and pitch accents have been predicted from text with decision trees and classifiers over part of speech, word class and information status [3], with conditional random fields [10], from lexical statistics in conversational speech [11], and with neural networks over word embeddings [12, 8]. Nielsen et al. [13] show that an acoustic pitch-accent detector gains from the context of the whole utterance, and that accenting every content word is already a strong baseline (82.2% accuracy on the Boston University Radio News Corpus), a point that our word-majority baseline echoes. Talman et al. [1] annotated LibriTTS for prominence and showed that fine-tuned BERT [14] outperformed recurrent and feature-based models on the task. Probing the layers of frozen BERT on these labels, Kakouros and O’Mahony [15] found in formation about prominence mostly in the middle layers, and the surprisal of words under LLMs correlates with prominence, although it helps speech synthesis only a little [16]. In speech, prominence has been detected both with supervised classifiers and with unsupervised methods, among them the wavelet-based method of Suni et al. [17], which was used to produce these labels, and 3PRO [18]. Self-supervised speech models [19, 20], speech recognisers [21] and speech-text LLMs such as Qwen2-Audio2 now provide representations of the signal that can be probed in the same way as text representations. Layer-wise probing has found suprasegmental information, including English phrasal pitch accents, mostly in the middle layers of such models [22], and prosodic structure, among it prominence, can be predicted from self-supervised representations of pitch, energy and timing alone [23]. Sentence stress can be detected reliably by a model built on Whisper [24], whereas speech language models, Qwen2-Audio among them, perform close to chance on the same task [25]. Audio-language models have also been found to encode prosodic and paralinguistic cues that are lost before the output or not used in their answers.3 Our speech results agree with these findings and extend the comparison to text models and prompted LLMs on the same labels.
Probing classifiers measure what information frozen representations make available [26]. A powerful probe may, however, learn the task itself [27], so we compare our classifiers with a control representation of random word vectors and base our claims about what representations contain on linear probes. In-context learning [28] with many examples has been shown to approach fine-tuning on some tasks [29]. Hu and Levy [30] show that LLMs’ answers to metalinguistic prompts can underestimate what is encoded in their probability distributions, and Hu and Frank [31] that demanding evaluation formats hide abilities of smaller models in particular. Similarly, linear probes on hidden states judge grammaticality better than the models’ own string probabilities [32], and probing reveals linguistic competence that prompting does not [33]. Similar gaps between what a model encodes and what it says have been found for truthfulness: linear probes and unsupervised methods recover from hidden states whether a statement is true more reliably than the model’s own answers [34, 35]. We add to this line of research with evidence from prosody, where the gap is large enough that prompted text LLMs do no better than a lexical baseline.
Dataset
The data are the prominence annotations of Talman et al. [1], which give a label to every word of LibriTTS, a corpus of read English derived from LibriSpeech audiobooks [36]. The labels were produced with the Wavelet Prosody Toolkit [17], which combines the fundamental frequency, energy and word durations of an utterance into one signal, analyses it with a continuous wavelet transform and assigns each word a score. The released data store this score next to the label. We find that the 3-way labels are fixed cut-offs of the score (0 below 0.4, 1 below 1.2, 2 above), without exception over 2,265,552 words, and that the score has been clipped at zero: null use is negative and 9.3% are exactly zero. We refer to the classes as 0 (non-prominent), 1 (prominent) and 2 (very prominent); the 2-way task merges
1 and 2. The annotations follow the LibriTTS splits (Table 1), which have no speakers in common. Text models are trained on train_360, the larger training set used in 2019. Since extracting speech representations for every frame of train_360 would be costly, speech models are trained on train_100, and to separate the effect of the input from that of the training data, we also train two text classifiers on it (§ 5.3). Models, layers and prompts are chosen on the development set, and we report word-level accuracy on the test set, including out punctuation. The three labels have nearly the same distribution in all four splits. As in 2019, the reported numbers leave out the last test utterance; this changes no result by more than 0.02 points.
| Split | LibriTTS subset | Utterances | Words | Speakers | 0 | 1 | 2 | Used for |
| train_360 | train-clean-360 | 116,263 | 2,076,289 | 904 | 48.3 | 27.4 | 24.2 | training (text) |
| train_100 | train-clean-100 | 33,042 | 570,592 | 245 | 48.1 | 27.3 | 24.6 | training (speech; matched text) |
| dev | dev-clean | 5,727 | 99,200 | 40 | 47.9 | 27.7 | 24.4 | model, layer and prompt selection |
| test | test-clean | 4,822 | 90,063 | 39 | 48.0 | 27.3 | 24.7 | evaluation |
Table 1. The data splits. Words: words with a prominence label (punctuation excluded). Labels: share of words in each 3-way class.
Reproducing the labels. Since the labels are a deterministic function of the recording and its word alignment, they can in principle be computed again. The original alignments are not available, so we use the public forced alignments of LibriTTS by Hayashi,4 made with the Montreal Forced Aligner [37], and run the toolkit with the configuration distributed for LibriTTS. On one test utterance for which the toolkit’s own alignment is distributed, both alignments give the original label for 15 of 15 words, and the two alignments change the scores by less than 0.05. On the whole test set, the reconstructed labels agree with the original ones on 91.5% (3-way) and 95.0% (2-way) of the words, and the scores correlate at . The remaining differences come from the alignments and from changes to the toolkit since 2019. The reconstruction serves as the upper reference for speech models: it is the labelling procedure itself, applied to our word boundaries.
Readers disagree. Some texts in LibriTTS are read by several speakers. On these 390 texts, excluding the LibriVox announcements, two readers agree on 77.0% of the 2-way and 63.8% of the 3-way labels. 189 test utterances also occur in the training or development data, read by another speaker. Every text model is 1.6–5.5 points less accurate on them than on the rest of the test set (3-way), so having seen the text during training does not help.
Experimental setup
Text models. The models are listed in Table 2. We fine-tune BERT-base with the original recipe [1]: Adam, learning rate , batch size 32, two epochs and the final-epoch model, with five seeds. BERT-large, ModernBERT-large [38] and XLM-R large [39] are fine-tuned with the same recipe, but with an attention mask and a learning rate of (three seeds each). One of the three 2-way runs of XLM-R large did not train and stayed at the level of the majority class on the development set; we leave such runs out of the averages. We probe 7 frozen encoders and 8 decoder LLMs from 0.8B to 30B parameters, representing each word by its first (encoders) or last (decoders) sub-word token. Decoders only see the left context, so we also give them the sentence twice and read the second copy (ECHO). A linear discriminant probe is trained on every layer and the best layer is chosen on the development set. On that layer we then train three classifiers on the frozen features, logistic regression, a multilayer perceptron and a bidirectional LSTM (BiLSTM), with three seeds each. As a control, the same BiLSTM is trained on random vectors, one fixed vector per word type, which carry no knowledge from any model.
| Model | Hugging Face identifier | Params | Revision |
| Encoders | |||
| BERT-base | google-bert/bert-base-uncased | 110M | 86b5e093 |
| BERT-base cased | google-bert/bert-base-cased | 109M | cd5ef92a |
| BERT-large | google-bert/bert-large-uncased | 336M | 6da4b6a2 |
| ModernBERT-base | answerdotai/ModernBERT-base | 150M | 8949b909 |
| ModernBERT-large | answerdotai/ModernBERT-large | 396M | 45bb4654 |
| XLM-R base | FacebookAI/xlm-roberta-base | 279M | e73636d4 |
| XLM-R large | FacebookAI/xlm-roberta-large | 561M | c23d21b0 |
| Decoder LLMs | |||
| Qwen3.5 0.8B | Qwen/Qwen3.5-0.8B-Base | 873M | dc7cdfe2 |
| Qwen3.5 2B | Qwen/Qwen3.5-2B-Base | 2.3B | b1485b2f |
| Qwen3.5 4B | Qwen/Qwen3.5-4B-Base | 4.7B | 1001bb4d |
| Qwen3.5 9B | Qwen/Qwen3.5-9B-Base | 9.7B | 68c46c4b |
| Qwen3.8 27B | Qwen/Qwen3.8-27B | 27.8B | 37561816 |
| OLMo 3 7B | allenai/Olmo-3-1025-7B | 7.3B | a81bae42 |
| OLMo 3 32B | allenai/Olmo-3-1125-32B | 32.2B | c2b61dae |
| Muse Glimmer 30B | meta-models/Muse-Glimmer-30B | 29.8B | cec94fcf |
| Speech | |||
| WavLM-large | microsoft/wavlm-large | 315M | c1423ed9 |
| HuBERT-large | facebook/hubert-large-ll60k | 315M | ff022d09 |
| Whisper large-v3 (encoder) | openai/whisper-large-v3 | 1.5B | 06f233fe |
| Qwen2-Audio 7B | Qwen/Qwen2-Audio-7B | 8.4B | dd844707 |
| Qwen2-Audio 7B Instruct | Qwen/Qwen2-Audio-7B-Instruct | 8.4B | 0a095220 |
Table 2. Models, with their Hugging Face identifiers, number of parameters (Whisper and Qwen2-Audio: the whole model) and the revisions used.
Prompting. GPT-4o, DeepSeek-V4-Pro and Mistral Large 3 receive the sentence and its numbered words and are asked for one label per word. We compare three instruction variants, three answer formats and 0, 1, 5 or 10 examples on 500 development utterances, and run the best setting of each model on the test set. The reasoning model GPT-5.5 is run on 200 test utterances with the setting chosen for GPT-4o. To see whether more examples help, we give the open model Qwen3.8 27B, which we also probe, between 0 and 500 examples (up to 146k tokens). It writes its answer word by word and we take the label it finds most probable, so that no answer needs to be parsed.
Speech models. Corpus words are matched to the aligned words, handling expanded abbreviations, unknown words and differences in tokenisation. We probe every layer of WavLM-large [20], HuBERT-large [19], the Whisper large-v3 encoder [21] and Qwen2-Audio 7B, both its audio encoder and its language model at the positions of the audio, representing each word by the mean of the frames inside its time span. The speech classifiers are trained on train_100; to compare text and speech with the same training data, we also train them on BERT-base and Muse Glimmer 30B on train_100. Qwen2-Audio 7B Instruct is prompted with the same prompt as above, once with and once without the recording. As a further reference, a linear classifier uses word timings alone (duration, number of phones, pauses and position in the utterance).
Comparing text and speech. Text and speech models are scored on the same words: those with a time span that starts within the first 30 s of the utterance, the input window of Whisper and Qwen2-Audio (89935 words, 99.9% of the test words). Differences are tested with McNemar’s test, and we report utterance-level bootstrap intervals.
Evaluation and compute. As in 2019, the primary metric is word-level accuracy. Due to class imbalance, we also report the macro-F1 scores and recall for each class for the primary systems. The open models are run on single NVIDIA GH200 GPUs (probing with bfloat16), and the commercial models are run via the Azure API. The runs used in this paper (including
controls not reported here) took a total of 58.5 GPU hours. The longest run was prompting Qwen3.8 27B with 500 examples on the test set, taking 8.1 hours.
Results
Table 3 gives an overview. Our reproduction of the 2019 model reaches 83.02 ± 0.13% (2-way) and 68.49 ± 0.10% (3-way), within 0.2 points of the published 83.2% and 68.6%. A baseline that gives each word its most frequent label in the training data already reaches 80.7% and 62.0%.
| System | 2-way | 3-way |
| Baselines | ||
| T Majority class | 52.0 | 48.0 |
| T Word majority (2019) | 80.7 | 62.0 |
| Fine-tuned | ||
| T BERT-base (2019 recipe) | 83.0 ± 0.1 | 68.5 ± 0.1 |
| T BERT-large | 83.1 ± 0.1 | 68.5 ± 0.1 |
| T ModernBERT-large | 83.2 ± 0.0 | 68.8 ± 0.1 |
| T XLM-R large | 83.3 ± 0.1 | 68.8 ± 0.2 |
| Frozen, linear probe | ||
| T BERT-base | 81.8 | 64.6 |
| T best text: Muse Glimmer 30B, echo | 82.6 | 66.9 |
| A best audio: Whisper large-v3 (encoder) | 88.2 | 78.7 |
| Frozen, BiLSTM classifier | ||
| T best text (3-way): Qwen3.5 0.8B, echo | 82.9 ± 0.0 | 68.0 ± 0.1 |
| T best text (2-way): ModernBERT-large | 83.0 ± 0.0 | 68.0 ± 0.1 |
| A best audio: Whisper large-v3 (encoder) | 91.0 ± 0.2 | 84.7 ± 0.2 |
| Prompting | ||
| T GPT-4o | 79.7 | 59.4 |
| T DeepSeek-V4-Pro | 79.2 | 60.5 |
| T Mistral Large 3 | 78.9 | 59.7 |
| T GPT-5.55 | 78.9 | 58.8 |
| T Qwen3.8 27B, 500 examples | 78.3 | 62.0 |
| A Qwen2-Audio, with recording | 51.0 | 39.2 |
| T Qwen2-Audio, text only | 51.6 | 43.1 |
| Audio references | ||
| A Duration only | 83.2 | 65.0 |
| A Labelling procedure | 94.9 | 91.4 |
Table 3. Test accuracy (%). Mean ± standard deviation over seeds where there are several. T: text input. A: audio input. Best layers and models are chosen on the development set. Bold: best text and best speech system in each column (baselines and references excluded). 5200 test utterances.
Text models
If more text and more parameters taught models more about prominence, the largest LLMs should come out ahead. Table 3 shows that they barely do. The linear probe of the best LLM, Muse Glimmer 30B, echo, reaches 82.6% and 66.9%. Scaling the Qwen models from 0.8B to 27B parameters raises the 3-way accuracy of the linear probe only from 65.7% to 66.7%, and that of the BiLSTM classifier not at all. With a BiLSTM classifier, the best frozen model reaches 83.0% and 68.0%. Finetuning larger and newer encoders does not move the ceiling either: BERT-large, ModernBERT-large and XLM-R large reach 68.5%, 68.8% and 68.8% on the 3-way task, against 68.49% for BERT-base.
Increasing the amount of training data does not raise the performance ceiling (Figure 1). The linear probe on BERT-base reaches 64.4% with 5,000 training utterances, and 64.6% when using all 116,263. Additionally, fine-tuning on 1/10 of train_360 yields 66.9%, whereas using all of it yields 68.5%. Therefore, what a text model can learn about prominence can be learned even with a small amount of data. Since even BiLSTM classifiers on random word vectors without a pre-trained model exhibit 66.6% performance, most of the performance achieved by text models can be learned solely from word identities and labels. Furthermore, although Talman et al. [1] suggested it could help, preceding discourse is not helpful. Providing BERT-base with one or three preceding sentences of a given chapter, or ModernBERT-large with three preceding sentences, changes the 3-way accuracy of both the probes and the classifiers by at most 0.1 points.

Figure 1. 3-way test accuracy of BERT-base with growing training data (subsets of train_360, train_100 and train_360).
Providing the entire sentence (ECHO) to the decoders adds 0.8–1.2 points on the 3-way task, suggesting that the following words matter, as can be expected from the accent at the end of a phrase.
The models also make the same mistakes. Among the 42 text systems (linear probes and classifiers for all text models and conditions, and therefore not independent systems), any two agree with each other on average for 94.0% of the words (2-way). However, each system agrees with the gold labels on only 82.0%. For 8.1% of the test words, all systems select the same incorrect label, which accounts for 45.2% of the average system’s 2-way errors (18.7% for 3-way). Therefore, while models with different architectures, sizes, and training data predict the same reading of the text, the gold labels differ from the predictions for the same words. This can be seen as an indicator that the ceiling is determined by the task, namely, how much of prominence is left to the reader.
Figure 2 shows the same pattern for all systems on the words used in Section 5.3. The text classifiers agree with each other on 84.4% of the words on average and with the gold labels on only 67.9%. The speech models agree with each other about as often (84.9%), but with the gold labels on 82.7%. The two kinds of model thus form two groups, which agree with each other on only 70.6% of the words.

Figure 2. How often two systems give a word the same 3-way label (%), on the 89,935 test words used for the text and speech comparison. Text and speech models are BiLSTM classifiers; white lines separate the gold labels, speech models, text models and prompted LLMs.
Probing and prompting
Since the contextual representations of LLMs contain information about prominence beyond word identity, one might expect the models to identify prominent words when asked. However, this is not the case. The best commercial models demonstrated an accuracy of 60.5% on the 3-way task (DeepSeek-V4-Pro) and 79.7% on the 2-way task (GPT-4o), which is lower than the word-majority baseline (62.0% and 80.7%). The format of the response was more important than the way instructions were expressed, and examples were not of much help. Adding 1 to 10 examples resulted in an improvement of only about 1 point on the 3-way task and no improvement on the 2-way task. The reasoning model GPT-5.5 also did not show particularly better performance, but these tests were tested using only 200 utterances, these results are for reference only.
Is information missing from the prompted model, or is the model unable to report that information? Table 4 compares the same model accessed in two ways. When probing with a linear classifier, Qwen3.8 27B achieved a performance of 66.7% in the 3-way task. When prompted with 500 examples, it achieved 62.0%, matching the performance of the word-majority baseline, while in the 2-way task, it achieved 78.3%, showing lower performance than the baseline. The linear probe can do little on its own. In the model’s embedding layer, which processes each word without context, it achieved 62.9%, showing performance close to the baseline. Therefore, the information the probe finds beyond the word’s identity lies in the model’s contextual representations, and the prompted model does not report this very information. Here, represented simply means that a linear classifier trained on labelled data can recover information from the hidden states. This does not imply that the model possesses a concept of prominence or can utilize it during generation. BiL-
| Model | Access | 2-way | 3-way |
| Qwen3.8 27B (T) | linear probe | 82.3 | 66.7 |
| Qwen3.8 27B (T) | BiLSTM classifier | 82.9 ± 0.1 | 67.8 ± 0.1 |
| Qwen3.8 27B (T) | prompted, no examples | 78.8 | 58.6 |
| Qwen3.8 27B (T) | prompted, 500 examples | 78.3 | 62.0 |
| Qwen2-Audio 7B (A) | linear probe (LM) | 86.4 | 74.4 |
| Qwen2-Audio 7B (A) | BiLSTM classifier (LM) | 87.3 ± 0.1 | 76.5 ± 0.1 |
| Qwen2-Audio 7B (A) | prompted (Instruct) | 51.0 | 39.2 |
Table 4. The same model probed and prompted (test accuracy, %). T: text input. A: audio input. Bold: best access for each model.
BiLSTM classifiers are less suitable for comparison because they can learn a significant portion of the task on their own. For random word vectors, BiLSTM classifiers achieve a performance of 66.6%, which is close to 67.8% achieved on Qwen3.8 27B (66.5% on its embedding layer). In the development subset, examples help only slowly, from 60.5% without examples to 63.5% when 500 examples are added, and they are of no help at all in the 2-way task (Figure 4).

Figure 4. Prompting Qwen3.8 27B with more examples (500 development utterances).
The speech-text LLM exhibits the same pattern in a stronger form. On Qwen2-Audio’s audio encoder, the BiLSTM classifier achieves 84.2%, and on its language model it achieves 76.5%, demonstrating far higher performance than all text models. However, when prompted with the recording, the instruction-tuned model only achieves 39.2% (when used without the recording, 43.1%), showing lower performance than the majority class. To verify if this reflects the forced response format, the model was allowed to respond freely on 20 test sentences. With the recording, the model repeats the numbered words without labels in 14 sentences, and provides labels in only 4 sentences (without it, 9 and 10, respectively). Even when labels are provided, it classifies almost all words as non-prominent (93 of 94 words when the recording is present, and 180 of 183 when absent). Therefore, the instruction-tuned model fails to perform the task properly, whether it responds freely or in the required format, even though its encoder represents prominence better than any text model.
Table 5 shows what the prompted models say instead: how often each system gives the highest level to content words, function words and the last word of the utterance. The BiLSTM classifier on Qwen3.8 27B reproduces the proportions of the gold labels almost exactly, while the prompted models give the highest level to the last word far too often: 81.4% of the time for GPT-4o and 77.1% for DeepSeek-V4-Pro, against 31.9% in the gold labels. This is not an effect of our instructions, which mention that prominence tends to fall towards the end of a phrase. With the minimal instructions, which say nothing about position, the three commercial models still give the highest level to 54.4–72.3% of the last words on the development subset (gold: 32.4%). With 500 examples, Qwen3.8 27B corrects much of this (45.5%), but it still falls 9.0 points behind the classifier on its representations on content words and 2.8 points on function words. The prompted models thus seem to answer with a rule of thumb about where the main accent of a sentence goes, while a classifier trained on the same model’s representations comes closer to how the sentences were actually read. Most of the difference is on content words, that is, on deciding which of the words that could be prominent actually are.
| System | Macro-F1 | Recall (%) | Labelled 2 (%) | |||||
| 0 | 1 | 2 | content | function | last | |||
| Gold labels | – | – | – | – | 42.2 | 9.2 | 31.9 | |
| T ModernBERT-large, BiLSTM classifier | 63.9 | 83.5 | 45.2 | 63.1 | 46.7 | 6.7 | 35.4 | |
| T Qwen3.8 27B, BiLSTM classifier | 63.8 | 83.2 | 48.4 | 59.2 | 43.1 | 6.0 | 31.7 | |
| P Qwen3.8 27B, prompted (500 ex.) | 55.9 | 84.2 | 35.4 | 48.3 | 44.2 | 3.4 | 45.5 | |
| P DeepSeek-V4-Pro, prompted (5 ex.) | 54.7 | 80.0 | 31.1 | 55.1 | 56.8 | 4.8 | 77.1 | |
| P GPT-4o, prompted (5 ex.) | 54.9 | 75.4 | 36.7 | 53.5 | 58.7 | 2.8 | 81.4 | |
| A Whisper large-v3, BiLSTM classifier | 83.0 | 90.4 | 72.4 | 86.6 | 43.4 | 9.3 | 32.7 |
Table 5. How each system labels the test set (3-way): macro-F1 and recall of each class, and the share of content words, function words and utterance-final words given the highest level (2). T: text classifier, P: prompted, A: speech classifier.
Accuracy hides some of the differences, as about half of the words are non-prominent and all systems classify most of the words correctly. Table 5 shows the macro-F1 scores and recall for each class. The classifier and the prompted model built on Qwen3.8 27B show a 7.8-point difference in macro-F1, larger than in accuracy. This difference appears in the prominent classes. The classifier detects 48.4% of words labelled 1 and 59.2% of words labelled 2, whereas the prompted model detects 35.4% and 48.3%. The confusion matrices in Figure 5 show the types of these errors. When asked directly, the model misclassifies 17.6% of very prominent words as non-prominent, whereas the classifier misclassifies 9.1%. Therefore, prompts not only blur the boundaries between levels but also miss prominent words entirely. On the other hand, the speech model rarely confuses the two extremes (1.3% of words labelled 2 are predicted as 0), and most errors occur between adjacent levels.

Figure 5. Confusion matrices (3-way, test set), normalised by row: the share of the words with each gold label that receive each predicted label (%).
Are the prompted models simply wrong about where the cut-offs between the levels lie? If so, better cut-offs on their own label probabilities should close part of the gap. We choose two cut-offs on the expected level, , on one half of the 500 development utterances and apply them to the other half, and vice versa. For Qwen3.8 27B with 500 examples, gives 63.0%, against 63.5% for the model’s own most probable labels and 62.4% for the word-majority baseline on the same words. The problem is therefore not the cut-offs but the order in which the model ranks the words. Its probability that a word is prominent separates prominent from non-prominent words with an AUC of 0.876, almost exactly as well as the share of prominent occurrences of each word in the training data (0.874), so ranking the words by the model’s own probabilities does not recover what the probe finds either.
Table 6 looks at the same question from the side of the lexical baseline. All text systems give most words the label they most often have in the training data, the classifiers on 72.4% of the words and the prompted models on 68.0–69.3%. The difference is in what happens when they depart from it. Where the classifier on Qwen3.8 27B departs from the lookup, it is right on 51.4% of the words and the lookup only on 30.8%, so the classifier uses the context to correct the lexical default. Where the prompted models depart from it, the lookup would have been right more often than the model (e.g., 44.2% against 36.0% for GPT-4o). What the prompted models report about prominence beyond the identity of the word thus seems to be mostly noise, while the departures of a classifier trained on its hidden states are not.
| System | Same as lookup | Departs | Right where it departs: system | Right where it departs: lookup | |
| T | ModernBERT-large, BiLSTM classifier | 71.7 | 28.3 | 51.8 | 30.7 |
| T | Qwen3.8 27B, BiLSTM classifier | 72.4 | 27.6 | 51.4 | 30.8 |
| P | Qwen3.8 27B, prompted (500 ex.) | 69.3 | 30.7 | 39.5 | 39.8 |
| P | DeepSeek-V4-Pro, prompted (5 ex.) | 68.8 | 31.2 | 37.1 | 42.1 |
| P | GPT-4o, prompted (5 ex.) | 68.0 | 32.0 | 36.0 | 44.2 |
Table 6. How often each system gives a word the label that the word most often has in the training data (the word-majority lookup), and, on the words where it does not, how often the system and the lookup are right (%, 3-way, test set).
Speech models
For the same words, the best speech model, the Whisper encoder-based BiLSTM classifier, achieved a performance of 84.6% in the 3-way task and 90.9% in the 2-way task. The best text models achieved performances of 68.1% and 83.1% (Table 8; all systems used in this comparison were selected on the development dataset). Therefore, the recordings add 16.5 points in the 3-way task and 7.8 points in the 2-way task. Since the labels were calculated from the same recordings, some of these benefits are inherent in the task itself (§ 7). These differences do not stem from the training data. When using the same train_100 data, the BiLSTM classifiers for BERT-base and Muse Glimmer 30B achieved performances of 67.1% and 67.5% (3-way), respectively, while Whisper reached 84.7%. All four speech models ultimately showed similar performance to each other (for the BiLSTM classifier, 83.9–84.7%), and the linear probes peaked in the middle and upper layers, specifically at a depth between 54% and 72% (Figure 3). Compared to the best text model overall, ModernBERT-large (fine-tuned) (68.8%), the best speech model is ahead by 15.9 points across the entire test set.

Figure 3. 3-way test accuracy of a linear probe on each layer, by relative depth (0: embeddings, 1: last layer).
Table 8: Accuracy (%) on the same 89935 test words, best systems of each kind. All differences from the best text model are significant (McNemar, ) unless marked ◦. Bold: best text and best speech system (references excluded).
| System | 2-way | 3-way | 95% CI (3-way) |
| A Labelling procedure | 95.0 | 91.5 | 91.3–91.7 |
| A Whisper large-v3 (encoder), BiLSTM classifier | 90.9 | 84.6 | 84.3–84.9 |
| A Whisper large-v3 (encoder), linear probe | 88.3 | 78.8 | 78.5–79.1 |
| A Qwen2-Audio 7B (LM), BiLSTM classifier | 87.4 | 76.7 | 76.4–77.0 |
| A Duration only | 83.2◦ | 65.0 | 64.7–65.4 |
| T Qwen3.5 0.8B, echo, BiLSTM classifier | 82.9◦ | 68.1 | 67.7–68.4 |
| T Muse Glimmer 30B, echo, BiLSTM classifier, train_100 | 82.8 | 67.5 | 67.2–67.9 |
| T Muse Glimmer 30B, echo, linear probe | 82.6 | 66.9 | 66.6–67.2 |
| T Qwen3.8 27B, prompted | 78.3 | 62.0 | 61.7–62.4 |
| A Qwen2-Audio 7B Instruct, prompted | 51.0 | 39.2 | 38.8–39.7 |
Table 8.
How much of the gap is timing alone? A classifier over word durations, phones and pauses reaches 83.2% on the 2-way task, as much as the best text model, but only 65.0% on the 3-way task, below it. Duration thus tells whether a word is prominent about as well as the text does, but not how prominent it is; for that, the models need pitch and loudness as well. Running the labelling procedure itself on our alignments gives 91.4% and 94.9%, which leaves 6.9 points (3-way) for speech models to learn, part of which is noise from the alignments.
The speech-text LLM loses some of what its encoder has learned. Within the encoder, the linear probe peaks at layer 22/32 and declines towards the encoder’s output; the language model receives this weaker signal, keeps it for a few layers and then slowly loses it (Figure 3). The best layer of the language model, 3/32, is near its input, and the BiLSTM classifier on it stays 7.7 points below that on the encoder.
Analysis
Word types
To gain a better understanding of where the recording helps, we break the comparison down by word type (Table 9). The gain is about twice as large for content words (22.9 points) as for function words (10.8 points), and it is largest on the last word of the utterance (23.4 points). Given and new content words gain about equally (22.5 and 22.7 points), which suggests that the speech models do not simply learn information structure better than the text models. Most importantly, on the words on which every text model makes the same mistake, the speech model is right on 58.9% of them, while it keeps 97.1% of the words that every text model gets right. The words that text cannot predict are thus largely the words where the recording decides.
Table 9: 3-way accuracy (%) of the best text and the best speech model on the same words, by word type and by whether the text models agree. : speech minus text.
| Words | Text | Audio | ||
| content words | 42,281 | 56.1 | 79.0 | +22.9 |
| function words | 47,654 | 78.7 | 89.6 | +10.8 |
| first word | 4,740 | 67.0 | 83.4 | +16.4 |
| medial | 80,394 | 69.1 | 85.2 | +16.1 |
| last word | 4,801 | 51.7 | 75.1 | +23.4 |
| new (content) | 29,629 | 56.4 | 79.1 | +22.7 |
| given (content) | 12,652 | 55.3 | 78.8 | +23.5 |
| unseen in training | 1,341 | 60.3 | 77.3 | +16.9 |
| seen 1–10 times | 4,385 | 58.9 | 80.1 | +21.2 |
| seen >1,000 times | 53,641 | 75.9 | 88.4 | +12.5 |
| every text model wrong | 9,699 | 0.0 | 58.9 | +58.9 |
| text models disagree | 44,625 | 57.4 | 80.2 | +22.8 |
| every text model right | 35,611 | 100.0 | 97.1 | −2.9 |
| all common words | 89,935 | 68.1 | 84.6 | +16.5 |
Table 9.
Table 10 gives an example. The text model places the accent on Paris, as one would expect when reading the sentence out of context, while the reader leaves Paris unaccented and stresses over and often instead. The speech model finds the same words as the reader, although it is too cautious with the highest level.
Table 10: One test sentence as labelled by the reader and as predicted by the best text model (ModernBERT-large) and the best speech model (Whisper large-v3), both with BiLSTM classifiers. Grey: not prominent (0); bold: prominent (1); large bold: highly prominent (2).
| Reader | Oh, yes, I go over to Paris often. |
| Text | Oh, yes, I go over to Paris often. |
| Speech | Oh, yes, I go over to Paris often. |
Table 10.
Readers of the same text
If the recording adds what the reader decides, speech models should predict the reading they hear rather than a reading of the text. Some test utterances have the same text as an utterance read by another speaker in the development set, which no model has been trained on. A text model receives the same input for both readings, so it can at best predict what the two readers have in common. Table 11 compares each system’s predictions with both readers. The speech models agree with the reader they hear on 80.3% (Whisper) and 81.3% (WavLM) of the words, but with the other reader only on 65.8% and 65.5%. This is close to how often the two readers agree with each other (64.2% on these pairs; 63.8% over all texts read by two other speakers). The text classifiers and the prompted models show no such preference: they agree with the two readers about equally often, and slightly more with the other reader. Although the sample is small, it does seem to support our interpretation of the gap between text and speech: the speech models do not only have better access to the same information, they follow the individual readings.
Table 11: 3-way agreement (%) of each system’s predictions with the reader of the test utterance (own) and with another speaker reading the same text in the development set (other), on 1,906 words in 143 test utterances. Classifiers are BiLSTMs. A: speech, T: text, P: prompted. : own minus other.
| System | Own reader | Other reader | |
| A Whisper large-v3 | 80.3 | 65.8 | +14.5 |
| A WavLM-large | 81.3 | 65.5 | +15.8 |
| T ModernBERT-large | 67.4 | 70.6 | −3.2 |
| P Qwen3.8 27B | 65.7 | 69.9 | −4.2 |
| P Qwen3.8 27B, 500 examples | 60.8 | 62.0 | −1.2 |
| P DeepSeek-V4-Pro, 5 examples | 59.1 | 58.7 | +0.4 |
| The two readers | 64.2 |
Table 11.
Table 7 shows an example. Speaker A places the main accent on I in the second clause, while Speaker B places it on must. A speech model listening to Speaker A’s reading places the accent on I, whereas a text model, which does not know who is reading, does not place the accent on I, like Speaker B. Neither reading is incorrect, and the text model can predict at best only one of the two. In fact, the text models match the reading of either speaker more often than the two speakers match each other (67.4% and 70.6% for ModernBERT-large, against 64.2%; Table 11). Therefore, the text model appears to predict something similar to the reading the speakers have in common, which is as close as text alone can get to any individual reading.
Table 7: One test sentence read by two speakers, typeset by prominence: speaker A (test set), the best speech model hearing speaker A, speaker B (development set) and the best text model, whose prediction is the same for both readings. Grey: not prominent (0); bold: prominent (1); large bold: highly prominent (2).
| Speaker A | Hilda’s face quivered, but she whispered Yes, I think it must have been. |
| Speech model, hearing A | Hilda’s face quivered, but she whispered Yes, I think it must have been. |
| Speaker B | Hilda’s face quivered, but she whispered Yes, I think it must have been. |
| Text model | Hilda’s face quivered, but she whispered Yes, I think it must have been. |
Table 7.
Discussion
Why do prompted LLMs fall short of what their representations contain? We see at least three possible reasons:
The labels define prominence acoustically, as computed from one reader’s voice, while the models’ notion of emphasis may follow the meaning of the sentence.
A graded three-level scale is hard to express in words, while a classifier learns the cut-offs directly from the labels.
Numbering the words and producing a fixed answer format add errors of their own.
The third reason is unlikely to explain much, since Qwen3.8 27B writes its answer under constrained decoding and still stays far below a linear probe on its representations. To gain a better understanding of the other two, we look at the predicted label distributions. It seems that the second reason is not supported by them either. The best prompted model predicts the three classes in roughly the right proportions (49.6, 21.1 and 29.3% against 48.0, 27.3 and 24.7% in the test set). However, it finds only 31.1% of the words labelled 1 and 55.1% of those labelled 2, against 50.2% and 58.7% for the best linear probe. The models
therefore choose the wrong words rather than the wrong amount of emphasis. Re-thresholding the model’s own label probabilities rules out the second reason more directly: better cut-offs do not help, because the model ranks the words no better than their frequency in the training data does. We believe that the first explanation is the most likely one: what the models report is a plausible reading of the sentence, built largely from what each word is usually like, not the reading in the recording. We have not tested this directly, but it would also explain why examples help so slowly, since each example shows one reader’s choices rather than a rule.
There is a parallel in human perception. Listeners asked to mark prominent words rely not only on the acoustic signal but also on their expectations about the words, among them how frequent the words are [40, 2], and marks from untrained listeners become reliable only when those of about ten listeners are pooled [41]. Explicit judgements of prominence thus draw on lexical expectations in humans too, and the prompted models’ answers, which follow the word-majority lookup, look more like such judgements than like the information a probe finds in their representations. This is in line with the distinction between what models encode and what they report on metalinguistic questions [30], and it suggests that asking an LLM about prosody tests its metalinguistic judgement rather than its knowledge of how sentences are read.
So, what is there in the text and what is there in the reading? The ceiling we observed is empirical and applies to the models, data, and contexts we tested. Preceding context, fine-tuning, and BiLSTM classifiers do not raise this limit, although this may partly reflect automatic labelling. Given this, our results present a simple distinction. Text has a large influence on determining prominence, particularly whether a word is prominent, and small encoders capture this well, just as large LLMs do. Where the preceding discourse calls for a contrast, LLMs can identify words to be emphasised from text [42]. However, in read audiobooks, much of the emphasis is not determined in this way. The degree of emphasis received by content words and where readers place the main accent of a phrase vary from reader to reader and reading to reading, and recordings complement this very aspect. As can be seen in Table 11, speech models follow individual readers, but text models do not. However, since speech models are still 6.9 points lower than the labelling procedure in the 3-way task, current speech representations do not even fully capture the acoustic part.
Under an independence assumption, the agreement between readers also bounds how high a text model could get at all. If each reading of a text is a draw from a distribution over the labels a word can receive given the text, two readers give a word the same label with probability . A text model, in contrast, can at best predict the most probable label, which is right with probability . Since , the agreement between readers bounds the best possible text-only accuracy from both sides: on average, it lies between 63.8% and 79.9% on the 3-way task and between 77.0% and 87.7% on the 2-way task (the upper bound by Jensen’s inequality). Resampling the 390 texts changes the 3-way bounds little (95% intervals 62.4–65.1% and 79.0–80.7%). The best text models (68.8% and 83.2%) lie within these bounds, in the lower half on the 3-way task. The same argument explains why the text models agree with either reader more often than the two readers agree with each other (Table 11): a model that predicts the most probable reading should. The bounds assume that readers are independent draws given the text, although the realisation of intonation varies systematically with the context and between speakers [43], and since the labels are computed automatically, part of the disagreement between readers may come from the labelling rather than from the readers.
These results have practical consequences. LLMs are increasingly asked to mark emphasis for speech synthesis or to pre-annotate speech data, while synthesised speech still conveys information status through intonation differently from human speakers [44]. Our results suggest that such a model, asked directly, will mostly reproduce a default placement of the main accent and miss many of the prominent words that its own representations encode. A small classifier trained on labelled data recovers more, on the model’s hidden states or even on the identities of the words alone, and at a fraction of the cost of prompting with hundreds of examples. For annotation, prominence labels proposed by a prompted LLM should therefore be treated as a weak prior rather than as a substitute for acoustic analysis. Since text can only determine a reading, not the reading, systems that need to match a particular speaker will also need the speaker’s own recordings or some other model of the speaker.
Some limitations should be kept in mind. Since the labels are produced automatically from the acoustics, speech models predict a function of their own input. Part of the gap between text and speech may therefore reflect how the labels were made rather than how listeners perceive prominence; perceptual labels on a subset of the test set would show how large this effect is. The labelling method itself was evaluated against manual accent annotation on the Boston University Radio News Corpus, where it was comparable with the best supervised methods [17], but it has not been validated against listeners on LibriTTS, and listeners may judge as prominent the words that carry an accent rather than the most salient ones acoustically [45]. Our word boundaries are not the ones the labels were computed on, but the reconstruction agrees with the original labels on 91.5% of the words, which suggests that the effect is small. Our data are read audiobooks, all published before 1923 and ranging from verse with a fixed metre to children’s stories, while the language models are trained on contemporary text. Prominence may also be easier to predict from text in read speech than in conversation. Our prompting results cover a limited set of instructions, answer formats and models, and the commercial models may change; a different elicitation, for example one that asks for a ranking of the words rather than labels, might bring out more of what the models encode. Finally, all our data are English, and languages in which prominence is more strongly tied to word order or morphology may leave less to the reader.
Conclusion
In this paper we have studied how much of prosodic prominence language models can find in text, whether they can report it, and what the speech signal adds. Text models from BERT to LLMs with 30B parameters converge on the same accuracy, at most 68.8% on the 3-way task, and classifiers trained on them share nearly half of their 2-way errors, which suggests that, for the models we tested, the ceiling lies in the task. With the prompts we tested, LLMs do no better than a lexical baseline. A linear probe nonetheless finds information about prominence in their contextual representations that they do not report. When asked directly, they place the main accent at the end of the sentence far more often than the readers do. Speech models add 16.5 points on the same words, mostly on content words and at the end of the utterance, and they are right on 58.9% of the words that no text model gets right. They also follow the individual reader: when two speakers read the same text, a speech model agrees with the one it hears far more often than with the other. Overall, our results highlight the difference between what a language model encodes and what it reports, and suggest that LLM-based synthesis front-ends and annotation pipelines should, where labelled data are available, read prominence from the models’ representations rather than ask for it in a prompt.
Several questions remain open for future work. Firstly, do the same results hold for perceptual labels and for conversational speech? Secondly, can speech-text LLMs be fine-tuned to report what their encoders contain? Finally, speaker-aware models could show how much of the remaining variation can be predicted once the reader is known.
Acknowledgements
The authors wish to acknowledge CSC – IT Center for Science, Finland, for computational resources. The second author gratefully acknowledges support from the Human Diversity consortium, under the Profi7 program (grant no. 352727) and the Centre of Excellence (grant no. 374221) by the Research Council of Finland. We thank Antti Suni for the Wavelet Prosody Toolkit and the prominence annotations of LibriTTS.
References
- [1]A. Talman, A. Suni, H. Celikkanat, S. Kakouros, J. Tiedemann, and M. Vainio, “Predicting prosodic prominence from text with pre-trained contextualized word representations,” in Proceedings of the 22nd Nordic Conference on Computational Linguistics (NoDaLiDa), Turku, Finland, 2019, pp. 281–290.arxiv.org/abs/1908.02262
- [2]D. R. Ladd and A. Arvaniti, “Prosodic prominence across languages,” Annual Review of Linguistics, vol. 9, pp. 171–193, 2023.
- [3]J. Hirschberg, “Pitch accent in context: Predicting intonational prominence from text,” Artificial Intelligence, vol. 63, no. 1–2, pp. 305–340, 1993.
- [4]M. Wagner and D. G. Watson, “Experimental and theoretical advances in prosody: A review,” Language and Cognitive Processes, vol. 25, no. 7–9, pp. 905–945, 2010.
- [5]D. Bolinger, “Accent is predictable (if you’re a mind-reader),” Language, vol. 48, no. 3, pp. 633–644, 1972.DOI
- [6]J. Yuan, J. M. Brenier, and D. Jurafsky, “Pitch accent prediction: Effects of genre and speaker,” in Proc. Interspeech, 2005, pp. 1409–1412.DOI
- [7]H. Li, L. Qu, J. Hu, and T. Li, “EME-TTS: Unlocking the emphasis and emotion link in speech synthesis,” in Proc. Interspeech, 2025, pp. 4368–4372.arxiv.org/abs/2507.12015
- [8]A. Rendel, R. Fernandez, R. Hoory, and B. Ramabhadran, “Using continuous lexical embeddings to improve symbolic-prosody prediction in a text-to-speech front-end,” in Proc. ICASSP, 2016, pp. 5655–5659.DOI
- [9]B. Stephenson, L. Besacier, L. Girin, and T. Hueber, “BERT, can HE predict contrastive focus? predicting and controlling prominence in neural TTS using a language model,” in Proc. Interspeech, 2022, pp. 3383–3387.arxiv.org/abs/2207.01718
- [10]M. L. Gregory and Y. Altun, “Using conditional random fields to predict pitch accents in conversational speech,” in Proceedings of ACL, 2004, pp. 677–683.DOI
- [11]A. Nenkova, J. Brenier, A. Kothari, S. Calhoun, L. Whitton, D. Beaver, and D. Jurafsky, “To memorize or to predict: Prominence labeling in conversational speech,” in Proceedings of NAACL-HLT, 2007, pp. 9–16.
- [12]S. Stehwien, N. T. Vu, and A. Schweitzer, “Effects of word embeddings on neural network-based pitch accent detection,” in Proc. Speech Prosody, 2018, pp. 719–723.arxiv.org/abs/1805.05237
- [13]E. Nielsen, M. Steedman, and S. Goldwater, “The role of context in neural pitch accent detection in English,” in Proceedings of EMNLP, 2020, pp. 7994–8000.arxiv.org/abs/2004.14846
- [14]J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of deep bidirectional transformers for language understanding,” in Proceedings of NAACL-HLT, 2019, pp. 4171–4186.arxiv.org/abs/1810.04805
- [15]S. Kakouros and J. O’Mahony, “What does BERT learn about prosody?” in Proceedings of the 20th International Congress of Phonetic Sciences (ICPhS), Prague, Czech Republic, 2023, pp. 1454–1458.arxiv.org/abs/2304.12706
- [16]S. Kakouros, J. Šimko, M. Vainio, and A. Suni, “Investigating the utility of surprisal from large language models for speech synthesis prosody,” in Proc. 12th ISCA Speech Synthesis Workshop (SSW), 2023, pp. 127–133.arxiv.org/abs/2306.09814
- [17]A. Suni, J. Šimko, D. Aalto, and M. Vainio, “Hierarchical representation and estimation of prosody using continuous wavelet transform,” Computer Speech & Language, vol. 45, pp. 123–136, 2017.DOI
- [18]S. Kakouros and O. Räsänen, “3PRO – an unsupervised method for the automatic detection of sentence prominence in speech,” Speech Communication, vol. 82, pp. 67–84, 2016.DOI
- [19]W.-N. Hsu, B. Bolte, Y.-H. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed, “HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 3451–3460, 2021.arxiv.org/abs/2106.07447
- [20]S. Chen, C. Wang, Y. Chen, Y. Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao, J. Wu, L. Zhou, S. Ren, Y. Qian, Y. Qian, J. Wu, M. Zeng, X. Yu, and F. Wei, “WavLM: Large-scale self-supervised pre-training for full stack speech processing,” IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, 2022.DOI
- [21]A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” in Proceedings of ICML, ser. Proceedings of Machine Learning Research, vol. 202, 2023, pp. 28 492–28 518.arxiv.org/abs/2212.04356
- [22]A. de la Fuente and D. Jurafsky, “A layer-wise analysis of Mandarin and English suprasegmentals in SSL speech models,” in Proc. Interspeech, 2024, pp. 1290–1294.arxiv.org/abs/2408.13678
- [23]S. Wallbridge, C. Minxhofer, C. Lai, and P. Bell, “Prosodic structure beyond lexical content: A study of self-supervised learning,” in Proc. Interspeech, 2025, pp. 4723–4727.arxiv.org/abs/2506.02584
- [24]I. Yosha, D. Shteyman, and Y. Adi, “WhiStress: Enriching transcriptions with sentence stress detection,” in Proc. Interspeech, 2025, pp. 4718–4722.arxiv.org/abs/2505.19103
- [25]I. Yosha, G. Maimon, and Y. Adi, “StressTest: Can YOUR speech LM handle the stress?” in Findings of the Association for Computational Linguistics: ACL 2026, 2026, pp. 1264–1285.
- [26]Y. Belinkov, “Probing classifiers: Promises, shortcomings, and advances,” Computational Linguistics, vol. 48, no. 1, pp. 207–219, 2022.arxiv.org/abs/2102.12452
- [27]J. Hewitt and P. Liang, “Designing and interpreting probes with control tasks,” in Proceedings of EMNLP-IJCNLP, 2019, pp. 2733–2743.arxiv.org/abs/1909.03368
- [28]T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. M. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. Amodei, “Language models are few-shot learners,” in Advances in Neural Information Processing Systems, vol. 33, 2020, pp. 1877–1901.arxiv.org/abs/2005.14165
- [29]R. Agarwal, A. Singh, L. Zhang, B. Bohnet, L. Rosias, S. Chan, B. Zhang, A. Anand, Z. Abbas, A. Nova, J. D. Co-Reyes, E. Chu, F. Behbahan, A. Faust, and H. Laroche, “Many-shot in-context learning,” in Advances in Neural Information Processing Systems, vol. 37, 2024, pp. 76 930–76 966.
- [30]J. Hu and R. Levy, “Prompting is not a substitute for probability measurements in large language models,” in Proceedings of EMNLP, 2023, pp. 5040–5060.arxiv.org/abs/2305.13264
- [31]J. Hu and M. C. Frank, “Auxiliary task demands mask the capabilities of smaller language models,” in Proceedings of the Conference on Language Modeling (COLM), 2024, pp. 1–17.arxiv.org/abs/2404.02418
- [32]Y. S. Wang, L. Qiu, Z. Wu, R. P. Levy, and Y. Kim, “Implicit representations of grammaticality in language models,” in Proceedings of ACL (Volume 1: Long Papers), 2026, pp. 15 035–15 055.
- [33]L. He, E. Nie, H. Schmidt, H. Schuetze, N. Mesgarani, and J. Brennan, “Large language models as neurolinguistic subjects: Discrepancy between performance and competence,” in Findings of the Association for Computational Linguistics: ACL 2025, 2025, pp. 19 284–19 302.arxiv.org/abs/2411.07533
- [34]A. Azaria and T. Mitchell, “The internal state of an LLM knows when it’s lying,” in Findings of EMNLP, 2023, pp. 967–976.arxiv.org/abs/2304.13734
- [35]C. Burns, H. Ye, D. Klein, and J. Steinhardt, “Discovering latent knowledge in language models without supervision,” in Proceedings of ICLR, 2023, pp. 1–28.arxiv.org/abs/2212.03827
- [36]H. Zen, V. Dang, R. Clark, Y. Zhang, R. J. Weiss, Y. Jia, Z. Chen, and Y. Wu, “LibriTTS: A corpus derived from LibriSpeech for text-to-speech,” in Proc. Interspeech, 2019, pp. 1526–1530.arxiv.org/abs/1904.02882
- [37]M. McAuliffe, M. Socolof, S. Mihuc, M. Wagner, and M. Sonderegger, “Montreal Forced Aligner: Trainable text-speech alignment using Kaldi,” in Proc. Interspeech, 2017, pp. 498–502.DOI
- [38]B. Warner, A. Chaffin, B. Clavié, O. Weller, O. Hallström, S. Taghadouini, A. Gallagher, R. Biswas, F. Ladhak, T. Aarsen, G. T. Adams, J. Howard, and I. Poli, “Smarter, better, faster, longer: A modern bidirectional encoder for fast, memory efficient, and long context finetuning and inference,” in Proceedings of ACL (Volume 1: Long Papers), 2025, pp. 2526–2547.arxiv.org/abs/2412.13663
- [39]A. Conneau, K. Khandelwal, N. Goyal, V. Chaudhary, G. Wenzek, F. Guzmán, E. Grave, M. Ott, L. Zettlemoyer, and V. Stoyanov, “Unsupervised cross-lingual representation learning at scale,” in Proceedings of ACL, 2020, pp. 8440–8451.arxiv.org/abs/1911.02116
- [40]J. Cole, Y. Mo, and M. Hasegawa-Johnson, “Signal-based and expectation-based factors in the perception of prosodic prominence,” Laboratory Phonology, vol. 1, no. 2, pp. 425–452, 2010.DOI
- [41]J. Cole, T. Mahrt, and J. Roy, “Crowd-sourcing prosodic annotation,” Computer Speech & Language, vol. 45, pp. 300–325, 2017.
- [42]A. Turetzky, A. Dekel, H. Aronowitz, R. Hoory, and Y. Adi, “Knowing what to stress: A discourse-conditioned text-to-speech benchmark,” in Proc. Interspeech, 2026, pp. 2932–2936.arxiv.org/abs/2604.10580
- [43]A. Arvaniti, A. Katsika, and N. Hu, “Variability, overlap, and cue trading in intonation,” Language, vol. 100, no. 2, pp. 265–307, 2024.
- [44]N. Hu, J. Kim, R. Orrico, S. Gryllia, and A. Arvaniti, “Can OpenAI’s TTS model convey information status using intonation like humans?” in Proc. Speech Prosody, 2024, pp. 32–36.DOI
- [45]R. Orrico, S. Gryllia, N. Hu, and A. Arvaniti, “The role of metrical structure in prominence perception,” in La voce nei media e nelle nuove tecnologie: produzione e percezione, ser. Studi AISV, V. De Iacovo, B. M. De Paolis, and D. Mereu, Eds. Milano: Officinaventuno, 2024, vol. 12, pp. 127–146.