betaFeedback

WorksComputer Science

FEND: Far-End Detection in Outbound Voice AI Calls

Abstract

When a conversational Voice AI agent places an outbound call, the far end may be a live person, a voicemail greeting, an interactive voice response (IVR) system, or increasingly another AI system answering on the callee's behalf. Standard Answering Machine Detection (AMD) reports only whether a human or a machine picked up, a design meant for human call centres, where one machine is as uninteresting as another. That single bit does not tell a Voice AI agent what to do, since leaving a voicemail message, navigating an IVR and holding a short exchange with another AI are three different actions and a two-valued label cannot distinguish them. In 1,000 randomly sampled calls from our own outbound deployment, a live person answered 48.1%, a voicemail system 44.6%, an AI 6.9% and an IVR 0.4%. Our proposed framework, FEND, reads the far end's audio and its streaming transcript and runs one detector for each automated class (voicemail, IVR, AI callee), with human the default when nothing fires, returning a single label within 20 seconds. On a public benchmark the voicemail components reach 100% accuracy without ever misclassifying a live human, where Asterisk's AMD misclassifies 30 to 66% of them. For the AI callee, acoustic features do not transfer across recording conditions, but speech timing does, separating AI from human at 0.952 ROC-AUC, because synthesised speech leaves far shorter gaps between words. On production traffic, separating a live person from any automated far end reaches F1 0.979, and only 2.9% of live people are given an automated label, the error that matters most for an agent acting on the result.

Introduction

An outbound call placed by a conversational Voice AI agent has to start by answering a question a human caller answers without thinking: who, or what, just picked up? Deployed practice reduces this to a single bit. Answering Machine Detection (AMD), the standard telephony service, reports whether a human or a machine answered [25, 14] (Twilio 2024a). That design comes from human call centres, where the point was to avoid spending a live agent’s time on a machine and one machine was as uninteresting as another.

When the caller is itself a voice AI agent the question changes. Talk time is cheap, the expensive mistakes are hanging up on a person and reading a script to an automated system, and the far end is no longer only a mailbox.

We therefore treat the far end as a four-class problem. HUMAN is a live person, and the agent should proceed. VOICEMAIL/AM is a mailbox greeting or an answering machine (AM), a device or carrier service that records on the callee’s behalf; the agent should leave its message and release the line. IVR is automation that speaks for an organisation, whether it offers selectable options, plays a queue announcement or reads a recorded business greeting; the agent should stay on the line. AI CALLEE is an automated system that answers for an individual and converses, responding to what the calling agent says and possibly handing the call to a person; the agent should introduce itself and hold a short exchange. Separating the last two by whom the automation speaks for, rather than by whether options are offered, is what makes the four classes disjoint. Merging any pair discards a decision the agent has to make, and merging AI CALLEE into VOICEMAIL/AM is the costliest, because it converts a conversation that could still reach the callee into a hang-up.

The classes are not equally served by existing work. We annotated 1,000 calls sampled at random from our own outbound deployment: a live person answered 48.1%, a voicemail system 44.6%, an AI system answering on the callee’s behalf 6.9%, and an IVR 0.4%. Consumer call screening is the reason the AI callee class exists at that scale, with handset assistants answering unknown numbers, asking who is calling and why, and deciding whether to pass the call on (Twilio, 2025).

Our contributions are the following.

  • A four-class formulation of far-end detection with disjoint classes, replacing the human-or-machine label that telephony AMD returns, and a system, FEND, that implements it.

  • An evaluation on a public benchmark and on 1,000 hand-annotated production calls, and against an open-source AMD implementation that can be replayed offline.

  • Evidence that the AI callee class is not separable from the far end’s spectrum but is separable from its speech timing, together with the cross-corpus procedure that distinguishes a real signal from a channel artifact.

To support replication we provide the class definitions, the corpus splits, the detector prompts, and a standalone reimplementation of Asterisk’s AMD module that reproduces its decisions call for call against Asterisk 20.6.0.

Related Work

Answering machine detection. Commercial platforms expose AMD as a configurable service returning a human-or-machine label. Twilio detects in about four seconds with a tunable speed and accuracy trade-off (Twilio, 2024a,b), Sinch signals when the beep arrives [25], Infobip reports the same two outcomes [14], and Telnyx adds a setting for Apple call screening, though its output is still one class per call [27]. Cisco [9] and CallFire (Thinakaran, 2008) describe pipelines over greeting length, the silence after it, and tone rejection, a Vonage project detects the beep with a learned audio model (Vonage (Nexmo) Community, 2019), and Asterisk’s app_amd is rule-based over silence and word-length thresholds [5, 18]. Research takes the same two-class view, from energy and cadence heuristics [12] to convolutional [3] and recurrent [2] models trained on contact-centre audio, and Saurav [23] detects voicemail in real time from fifteen temporal speech-activity features at 96.1% accuracy over 77,000 production calls. A vendor evaluation of voicemail detection in outbound agents reports the same limitation (Zammit, 2026). None of this work distinguishes the kinds of machine, which is the distinction a voice AI agent needs.

Detecting an AI on the line.

The natural instrument for the AI callee class is the anti-spoofing literature, where ASVspoof 5 is the current large-scale benchmark (Wang et al., 2025). Two recent results bear on our setting: Sharma and Wang [24] find most of eight pretrained detectors near or below chance against current synthesis under post-processing, and Chandra et al. [7] report roughly a 48% drop in audio detector area under the curve on in-the-wild deepfakes. Our results agree with both. The other place to look is timing. Consumer call screening is the immediate source of this class and vendors have begun documenting how it interacts with AMD (Twilio, 2025), while work on turn-taking and endpointing [26, 22, 15] treats conversational timing as a control signal. We use timing instead as evidence about who is speaking, which to our knowledge has not been done for far-end classification.

Methodology

A far-end detector observes one channel of a live call, the side the callee is on, as a stream of audio frames and of partial transcripts produced by automatic speech recognition (ASR). At every ASR update it may emit a label, and the deployment commits to whichever label stands at a decision deadline WW seconds after the call is answered. HUMAN is the default: a detector must produce positive evidence to assign any other class, so a live person is mislabelled only when some detector fires in error.

No single model sees all four classes, because the evidence that identifies them is of different kinds and arrives at different times. FEND therefore runs one independent detector per kind of evidence over two taps on the same call, the raw audio frames and the streaming transcript, each emitting a binary firing for a single class Figure 1. Each keeps running after it has fired, because the far end can change during a call when a screener hands over to a person or a greeting ends in a beep. Two of them are tool calls on the large language model (LLM) that is already running the dialogue, so they cost no extra model call.

Figure 1. Detection architecture. Four detectors read the raw audio frames and the streaming transcript and run concurrently, each able to fire one class. Firings accumulate over a window of WW seconds and the highest-priority class that has fired wins, with HUMAN the default when nothing fires. § 5.4 sweeps WW and the ordering.

Tone detector. Looks for the record-now beep that follows a mailbox greeting, using a short-time Fourier transform over the incoming audio frames (Appendix E). It is the only detector that does not depend on the ASR, so it is also the only one that still works when the greeting is in a language the ASR handles poorly.

Keyphrase detector. A multilingual list of voicemail phrases, such as “after the tone” or “please leave your name”, matched against the normalised running transcript. It is precise, cheap, and trivially extended to a new language, but it only fires on wordings someone has already added to the list (Appendix F).

Voicemail LLM detector. An LLM tool call that fires when the transcript shows an automated mailbox system answering for the callee (Appendix G).

IVR LLM detector. A second tool with its own prompt, which fires on automation speaking for an organisation, such as menus, queue announcements and recorded business greetings (Appendix H).

AI callee detector. Measures how the far end speaks rather than what it says, since an assistant answering for an individual replies to whatever the voice AI agent says and cannot be recognised from a fixed script. The ASR already reports when each word starts and ends, and from the far end’s first turn we summarise the pauses between words and how fast it speaks. A person pauses to think and to breathe, while a synthesised turn runs on. The detector is a logistic-regression classifier over seven such timing summaries, with its threshold set on the training half at a 5% false-positive budget against live humans (Appendix K).

Combining the detectors. Detectors run concurrently and more than one can fire on the same call, so a combination rule is needed. Two choices define it: the deadline WW, and what to do when two detectors fire inside it. We accumulate firings over the window and return the highest-priority class that has fired by the end, ordered voicemail, IVR, AI callee, with human as the default. Priority is by how conclusive the evidence is, so the class whose evidence is hardest to produce by accident wins, which is not the same as the class that fires first. We treat both WW and the ordering as parameters and sweep them in § 5.4 instead of fixing them by assumption.

Experiments

Datasets

Telephony-AMD (public). 400 labelled far-end clips across English, Spanish, French and German, with four classes: HUMAN, VOICEMAIL, ANSWERING MACHINE and IVR ([1]). We split it once, stratified by class and language, into a 119-clip tuning half and a 281-clip evaluation half. Prompts and thresholds are fixed on the tuning half before the evaluation half is scored.

In-house Voice AI calls. 1,000 outbound calls sampled uniformly at random from our deployment, annotated manually with far-end ground truth. Because the draw is random, the counts estimate the mix the deployment encounters: HUMAN 481 (48.1%), voicemail or answering machine 446 (44.6%), AI callee 69 (6.9%) and IVR 4 (0.4%). Half the sample fits the one detector trained in-house and all in-house results are reported on the other half. These calls contain customer content and are not released. Two forces are enlarging the AI callee share. Handset vendors are rolling out call screening to consumers, placing an assistant in front of a large share of personal numbers, and disclosure rules such as EU AI Act Article 50 ([11]) and California’s B.O.T. Act [6] require automated callers to identify themselves, which turns part of the class into one that announces itself.

Corpora for the acoustic analysis. The acoustic test of § 5.3 needs speech whose origin we know, so we assemble four sets and pass them all through the same channel simulation and the same fixed-length window. Two are human speech from independent sources, recorded bank calls (Wu et al., 2020) and role-played call-centre dialogues across fourteen English accents [4]. One is public synthesised speech [8], the positive class. The fourth is a public robocall corpus [20], which acts as a control: robocalls are mostly pre-recorded human speech, so they show whether a detector is responding to synthesis or merely to audio that is not a live two-way call (Appendix § 10).

Models

Two different recognisers are involved, one per corpus. On the public Telephony-AMD clips we run Whisper small [21] in streaming mode with SimulStreaming [17, 16], so that the public numbers are reproducible. The in-house calls are transcribed by the commercial ASR the deployment already runs, which we cannot disclose. The cadence features of § 5.3 are read from its word timings, so the timing result is tied to that recogniser. Both LLM detectors use GPT-4o [19] and the prompts are frozen before evaluation (Appendices § 15 and § 16).

Metrics and Evaluation

We report standard per-class precision, recall and F1, and for the ranking detectors the area under the receiver operating characteristic curve (ROC-AUC) together with recall at a fixed budget on human false positives. The most costly error is giving a live person an automated label, because the agent then ends a conversation with a real human. We therefore treat the human false-positive rate (HFPR), the fraction of true HUMAN far ends given any other label, as a constraint and not as one metric among several. For the ranking detectors we fix an HFPR budget of 1–5% and report the recall reached inside it, because a detector with a high ROC-AUC but almost no recall inside that budget cannot be deployed.

One evaluation choice is specific to the acoustic study. A measurement that separates two corpora of human speech from each other cannot be evidence about synthesis, so we calibrate each candidate on human data alone before any human-versus-machine comparison. For values SS on one human corpus and RR on a second, independent one, we admit the measurement only if the median difference, scaled by the interquartile range (IQR) of the reference corpus, stays small:

∣g∣≤1,g=med⁡(S)−med⁡(R)IQR⁡(R).|g| \le1,\qquad g = \frac{\operatorname{med}(S)-\operatorname{med}(R)}{\operatorname{IQR}(R)}.

Four of our five candidate measurements fail this null calibration (Appendix § 17).

The remaining protocol follows the datasets. On Telephony-AMD, thresholds and prompts are frozen on the 119-clip tuning half and the 281-clip evaluation half is scored once, with positive = VOICEMAIL+AM and IVR held out for the voicemail experiment and positive = IVR for the IVR experiment. The baseline is our offline port of Asterisk app_amd at shipped defaults and at the best of 1,152 grid-searched configurations (Appendix § 9). The timing and four-way experiments use the in-house calls, split by source call so that no call appears in both halves, fitted on one half and reported on the other at the deployment’s own class distribution, with 200 resplits giving the reported intervals.

Results

Voicemail and Answering Machine

Table 1 reports the three voicemail detectors and their combination on both splits of Telephony-AMD, with Asterisk for comparison. The key property holds on both splits: no detector fires on any human clip, so precision is 1.000 throughout. On the evaluation half the combination reaches 100% accuracy at 1.000 recall. The tone detector is stronger on answering machines (96.9% against 81.9% on voicemail greetings), the LLM is the reverse (100.0% against 98.5%), and the keyphrase list is by far the most conservative.

The open-source baseline. We ported Asterisk app_amd to an offline harness and validated it clip-for-clip against Asterisk 20.6.0. With shipped defaults it labels 65.8% of live humans as machines. Tuned by grid search over 1,152 configurations it still labels 29.8%, and it fires on every IVR clip (Appendix § 9). The thresholds are set for a far end that answers with a short “Hello?”, so a live person who says more than a word or two runs past the greeting-length and word-count limits and is declared a machine.

ComponentVMAMHumIVRAccRec
tuning half (nn=90 scored, IVR held out)
Tone (beep)70.0100.00.00.00.9000.842
Keyphrase50.025.90.00.00.6110.386
LLM (voicemail)96.7100.00.00.00.9890.982
Fusion96.7100.00.00.00.9890.982
evaluation half (nn=215 scored, IVR held out)
Tone (beep)81.996.90.00.00.9300.891
Keyphrase47.236.90.00.00.6330.423
LLM (voicemail)100.098.50.00.00.9950.993
Fusion100.0100.00.00.01.0001.000
Asterisk app_amd, all 400 clips
shipped defaults93.198.965.892.40.7310.959
tuned (5-fold CV)100.098.929.8100.00.8860.995

caption: tbl:1: Voicemail components on Telephony-AMD. Binary metrics take positive = VM+AM with IVR held out. Precision is 1.000 for every component row: none fires on any human clip. The app_amd human fire rate is the column that matters, and it is why a detector built for human call centres cannot be reused here.

:::

IVR

On the held-out split the IVR detector reaches 93.8% recall at 100% precision in two of three runs, and never fires on a human clip on either split. Two choices of definition produced most of that number. First, an IVR is not only a menu: queue announcements such as “your call will be answered in the order it was received” and recorded business-hours greetings are automated far ends that offer no option to select, and admitting them raised recall from 60.7% to 96.4% on the tuning half. Second, breadth of that kind has a cost once a fourth class exists, because an automated far end that is not a mailbox also describes an assistant answering for a person. We therefore scope the class by whom the automation speaks for: an IVR answers for an organisation. With that scoping the detector fires on none of the AI callee calls in our production sample, and the public benchmark is unchanged (Appendix § 16).

AI Callee

This is the class where content does not help. Acoustic evidence does not carry it: of five measurements of the far-end signal, four fail the null calibration of § 4 because they track how much of the window contains speech rather than who produced it, and the survivor reaches only 0.624 ROC-AUC with zero recall at a 1% budget. Repeating the comparison inside a single recording channel appears to reach 0.858, but one high-frequency energy measurement reaches 0.866 alone and gives 0.512 on a voicemail control. We do not report it as a detection result (Appendix § 17).

Cadence featureAIHumanAUC
Articulation rate (w s−1^{-1})3.682.510.844
Inter-word gap, CV3.501.760.834
Inter-word gap, p90 (s)0.0000.1380.687
Word duration, CV0.3730.4470.706

caption: tbl:2: Far-end timing on matched production calls. The cadence separates recorded or synthesised speech from live conversation.

:::

Speech timing does carry it. Word timings survive any codec, and using only how the far end’s first turn is paced, a model reaches ROC-AUC 0.952 [0.902, 0.990] on held-out calls with recall 0.774 at a 5% budget (Table 2). A person pauses to think and to breathe, opening gaps between words, while a synthesised turn runs on and is delivered faster. Table 2 lists the four most informative of the seven features the model uses, all read from our deployment’s ASR word timings. Two controls bound the claim: the same model separates AI from voicemail at 0.895, so it is not merely detecting pre-recorded audio, and the voicemail control sits at 0.707. Only 62 of the 481 human far ends open with at least four timed words, so the comparison is against the more talkative humans, and the positives appear to come from a single screening product. In 17 of the 69 AI callee calls in the sample the far end answers our agent by name, replying to its self-introduction with a phrase built around it, which requires processing speech online and is something no mailbox greeting can do.

All Four Classes Together

Table 3 reports FEND end to end on the held-out half of the production calls, where all four classes are labelled and the distribution is the unbalanced one the deployment sees. The decision that gates the agent is solved. Separating a live human from any automated far end reaches F1 0.979, at precision 0.987 and recall 0.971: of the 241 live humans, 234 keep the HUMAN label and 7 are given an automated one. Overall four-way accuracy is 0.938. What remains hard is telling the automations apart. AI callee recall is 0.457 at precision 0.762, so when that detector survives arbitration it is usually right and its errors are missed detections,

ClassnPrecRecF1FP into
HUMAN2410.9870.9710.979IVR 1, AI 1
voicemail/AM2230.9120.9820.946AI 18, H 3
IVR20.3330.5000.400n too small
AI callee350.7620.4570.571VM 2, H 3
HUMAN vs rest2410.9870.9710.9797 humans

caption: tbl:3: Per-class results on the held-out half (n=501n=501) of the in-house calls, at the deployment’s own class distribution, under a 20 s detection window. FP into names the true classes most often given this label. Overall accuracy 0.938. Only the AI callee component is fitted on in-house data, on the other half. The IVR row rests on two calls and is reported for completeness only.

not false alarms, with 18 of 35 AI calls assigned to voicemail. The confusion runs one way, and it runs in the expensive direction: treating an AI callee as voicemail throws away a conversation that could still reach the callee, while the reverse only wastes a message.

The window governs the human decision; one ordering relation governs the rest. On the public evaluation half, accuracy rises from 94.3% at a 5 second window to 98.2% at 15 seconds and is flat after, so a deployment that commits within the three to five seconds common in practice is deciding before the evidence has arrived. Sweeping the priority order on the in-house half, the human-versus-automated decision is invariant: F1 0.979, with exactly 7 of 241 live humans mislabelled under every ordering. Among the automated classes one relation carries the result. Ranking voicemail above AI callee is worth 14 points of four-way accuracy, 0.938 against 0.796, because inverting it sends 88 voicemail calls to the AI callee label rather than 18 the other way, and AI callee precision falls from 0.762 to 0.260. That is the relation conclusiveness prescribes: a record-now beep or a menu prompt is near-decisive, while timing evidence fits three of the four classes. The position of IVR moves accuracy by 0.2 points (Appendix § 18).

Latency, Cost and Deployment Impact

Latency and cost. Table 4 reports time from call answer to firing on the production calls. These are end-to-end decision times and not compute times: every detector that reads text waits for the ASR, and the tone detector waits for the greeting to reach its beep. Compute is negligible for three of the four, and only the LLM has a real per-call cost, which is prompt tokens on a call the model is already making, since it is a tool on the model already running the dialogue. The ordering is informative. The AI

DetectorMed.p90Cost
Cadence (AI callee)2.02 s6.98 snegligible
Keyphrase3.24 s14.70 snegligible
Tone (beep)8.86 s16.25 snegligible
LLM (voicemail)8.91 s13.07 s604+41 tok

caption: tbl:4: Time from call answer to firing on the 1,000 in-house calls, conditioned on firing within the 20 s window, with per-call cost. Token counts are median prompt and completion.

callee detector is the fastest at a median of 2.02 seconds because it needs only a few timed words, while the evidence that a mailbox is speaking takes three to four times longer. The class our arbitration ranks last is therefore the one whose evidence arrives first, which is why priority is by how conclusive the evidence is and not by when it appears.

Deployment considerations. On the in-house mix, 51.9% of answered calls are not a live person, so more than half of all outbound attempts are decided by this component and not by the dialogue that follows. Of those, 6.9 points are AI callees, a class the binary label cannot express: under a human-or-machine policy each is treated as a mailbox and the call ends, discarding the one automated far end that can still route the agent to the person it called. In the other direction, HFPR bounds the damage the system can do. At 2.9%, 7 of the 241 live people in the held-out half are given an automated label. This is what makes a four-class detector usable on live traffic.

Conclusion

We treated the far end of an outbound Voice AI call as a four-class detection problem, evaluated on a public benchmark and on 1,000 hand-annotated production calls. VOICEMAIL/AM is solved from content, at 100% accuracy on a held-out public split with no false positives on humans, where the open-source baseline mislabels 30 to 66% of live humans. IVR becomes separable once the class is defined by whom the automation speaks for. The AI CALLEE is the open case, and the class AMD has no label for. Its spectrum does not identify it, since measurements that appear to work either fail a human-versus-human calibration or track the recording channel; its speech timing does, at 0.952 ROC-AUC. End to end, the decision that gates the agent reaches F1 0.979, while telling the automations apart remains the harder half. The fastest-growing far end is the one that must be detected from how it speaks, not from what it says.

Limitations

The in-house class mix is one deployment’s. The 1,000 calls come from one outbound programme, so the proportions describe its callee list, calling hours and country mix and not outbound Voice AI in general. The IVR row of Table 3 therefore rests on two calls and is reported only for completeness, and the AI callee rate of 6.9% should be expected to move quickly across markets, since consumer call screening is switched on by handset vendors and not adopted call by call.

Every AI callee we observe is a screener. All AI callee calls in our sample are the same subtype, and a gatekeeping assistant with a scripted opening. We have no measured instance of an AI receptionist, personal assistant or third-party agent conducting a call, so the claim that the class will grow beyond screening rests on where the technology is going and not on our data, and the timing model establishes only that one product can be told apart.

The acoustic result is negative, and tested only against read-prompt speech. Our only usable synthetic arm is VoiceBench, which is read from prompts. The arm that would have tested conversational synthesis over a live carrier failed a channel check and was excluded, and the window normalisation retains 92% of one human corpus but only 48% of the synthetic one, leaving a selection difference between the classes compared. After calibration one metric survives, so the 0.624 AUC rests on a single statistic. No corpus of genuinely far-end agent-answered calls exists, and building one is the necessary next step.

The timing result depends on the deployment’s ASR and on a talkative minority. The 0.952 AUC is measured on 67 AI and 62 human calls, the 13% of people whose opening turn carries at least four timed words. The timings come from the deployment’s commercial ASR, so a recogniser with different word-boundary behaviour could align synthesised speech differently, and we have not reproduced the result on an open ASR.

The comparisons are uneven. The public corpus is short clips and probably under-represents long greetings and human pickup after a greeting, which is why the four-class evaluation runs on in-house traffic that is not public. Commercial AMD runs inside the carrier media path and cannot be replayed, so those numbers stay published operating points. The closest learned prior work [2] releases no checkpoint, so we compare against a rule-based system only, and the LLM detectors use a closed model whose pre-training data we cannot audit (Appendix § 11).

Ethical Considerations

The in-house calls contain customer content and are not released, all examples are anonymised, and no personal data is reproduced. Outbound messaging should still respect applicable consent, disclosure and do-not-call rules.

Deciding whether the other party is human. The AI callee class asks a system to judge whether a person is on the line, and the error is asymmetric morally as well as operationally, since treating a human as a machine can mean talking over them, hanging up on them, or sending them down a degraded path. That is why we report the operating point under a budget on human false positives and treat a detector with zero recall at that budget as undeployable. The inference is also adversarially exposed, because an operator who wants to avoid detection can change pacing at little cost. We describe the features anyway, since a closed detector whose failure modes nobody has measured is worse for the people on the receiving end.

Detection is not a substitute for disclosure. EU AI Act Article 50 [11] and California’s B.O.T. Act [6] require automated callers to identify themselves. Inferring what regulation says should be volunteered risks making non-disclosure feel normal. We regard detection as a defensive complement to disclosure, not a reason for a calling agent to disclose less.

Treating callers differently by class. Where the inferred class is wrong, the difference in treatment lands on a real person, and the far ends most likely to be misread, such as accented speech or poor channels, are not evenly distributed across callees. Our public evaluation covers four languages and our human reference corpora are two call-centre populations, so we cannot say how error rates vary across speaker groups, and that measurement is a prerequisite for deploying class-based policy at scale. Relatedly, what an agent should do on recognising another agent is a question about whose interests the call serves, which we leave as a configured policy.

Asterisk app_amd Port

The baseline of § 5.1 is a line-for-line port of two Asterisk sources: the isAnsweringMachine state machine of apps/app_amd.c, which is a threshold cascade over word and silence durations, and the silence detector of main/dsp.c, which declares a frame silent when its mean absolute sample value falls below silence threshold and accumulates consecutive silent frames. We preserve the order in which app_amd tests its thresholds, its initial IN_WORD state (so the first word of a greeting is not counted), and the fact that voice duration accumulates across the whole call while consecutive voice duration resets on any silence run reaching between words silence.

Audio is resampled to 8 kHz mono signed-linear and framed at 20 ms (160 samples), matching an RTP-fed channel. A recording ends but a connected call does not, so clips shorter than total analysis time are padded with digital silence, which is what a channel that stays up delivers; 79 of 400 clips are affected at the shipped defaults. app_amd returns MACHINE, HUMAN or NOTSURE, and a deployment must map the last of these to an action; we score both mappings and report the conservative one (NOTSURE → human, which protects live callers). It changes nothing material here, because only one test clip returns NOTSURE at the defaults and none does under the tuned configuration.

The grid search covers initial silence ∈{1.5,2.5,4} s\in\{1.5, 2.5, 4\}\,\mathrm{s}, greeting ∈{1.5,2.5,4,6} s\in\{1.5, 2.5, 4, 6\}\,\mathrm{s}, after greeting silence ∈{0.4,0.8,1.5,2.5} s\in\{0.4, 0.8, 1.5, 2.5\}\,\mathrm{s}, total analysis time ∈{5,10} s\in\{5, 10\}\,\mathrm{s}, maximum number of words ∈{2,3,5,8}\in\{2, 3, 5, 8\} and silence threshold ∈{128,256,512}\in\{128, 256, 512\}: 1,152 configurations. The selected configuration raises every threshold well above Asterisk’s defaults (4 s initial silence, 4 s greeting, 2.5 s after-greeting silence, 10 s analysis, 5 words), which is the search telling us the defaults are tuned for a far end that answers with a short “Hello?” rather than for the greeting lengths in this corpus.

Dataset Composition

Table 5 gives the composition of the Telephony-AMD test split used throughout § 5.

HUMANVMAMIVR
Clips1141029292

Table 5. Telephony-AMD test split (400 clips): class composition (top) and ASR-detected language composition (bottom).

The internal sets are: a beep golden set of 293 beep-positive clips (tone ~350–1000 Hz) and 74 beep-negative conversations; a beep-free set of ~65,000 clips including hold music and filler speech, used only for false-alarm validation; and a random sample of 1,000 outbound in-house Voice AI calls drawn over a fixed period, of which 478 showed no AMD activity and 464 were detected by at least one detector. These cannot be released, for compliance and data-privacy reasons, as with the proprietary corpora of prior work ([2, 3]); the Telephony-AMD evaluation is public and repeatable.

Data Contamination

The LLM detector is a general-purpose model prompted zero-shot and is not fine-tuned on any AMD corpus, so contamination could only take the form of Telephony-AMD audio or transcripts having entered pre-training. We cannot audit a closed model’s corpus, but three observations bound the concern. First, the detector is not asked to recall a benchmark label: it sees a streaming transcript and decides whether to emit a tool call, and is scored on whether and when it fires inside a live call, not on a held-out label. Second, its behaviour on Telephony-AMD matches its behaviour on our in-house Voice AI traffic, which post-dates the model and is not public, in both it is the sole signal on the same category of far end (AI screeners, conversational greetings with no beep). Third, the two detectors carrying most of the recall, beep and keyphrase, involve no learned model at all, so the headline OR-fusion result does not rest on the LLM path.

Behaviour on IVR

IVR is a third far-end behaviour, not a mislabelled voicemail: the correct agent response is to stay on the line and navigate the menu, never to leave a message and terminate. On the 92 IVR clips of the Telephony-AMD test split the beep detector fires on 0 (0.0%), since menus carry no record-now beep, and the keyphrase detector on 0 (0.0%), since no voicemail phrase matches a menu prompt. The LLM detector fires on 7 (7.6%). Three of those seven are genuinely ambiguous far ends, two hold queues that themselves offer a voicemail fallback (“Votre appel est en attente. . . Pour laisser un message. . . ”) and one automated business-hours announcement, and the remaining four fired on short ASR fragments (“Your call. . . ”) and are prompt failures we count against the detector. A spurious IVR firing does not by itself end the call, but it does put the agent on the wrong track. Tightening the prompt against fragmentary transcripts is future work.

Beep Detector Details

What a beep looks like. From manual inspection of collected beeps, the tone is close to a single sinusoid: frequency 350–2000 Hz, duration 0.5–2.0 s, one dominant spectral peak with few harmonics, and energy in a narrow band. The hard cases are a beep masked by speech or environmental noise, carrier frequency jitter, variable duration, codec artifacts, and tonal hold-music that mimics a beep.

Four conditions. A beep is declared only when all four hold on the same frequency bin, each closing one way a non-beep could pass:

  1. Dominant peak: the top peak exceeds the second by ΔdB=20log⁡10(M1)−20log⁡10(M2)≥20\Delta\mathrm{dB}=20\log_{10}(M_1)-20\log_{10}(M_2)\ge20 dB. A beep is one pure tone, whereas speech is broadband.

  1. Energy isolation: the mean energy just below (ElowE_{\mathrm{low}}) and just above (EhighE_{\mathrm{high}}) the peak are both under threshold. It rejects broadband noise and music, whose energy spreads across neighbouring bins.

  1. Frequency floor: the peak sits above 350 Hz. It rejects low-frequency hum, voiced filler, and DTMF tones.

  1. Streak: the peak persists for streak required = 4 consecutive frames. It rejects transient tonal blips.

The streaming pipeline is: PCM → float32 normalization → windowed STFT (Hann window [13], computed with the FFT [10]) → peak detection (top bins, neighbour merge, dominance check) → energy/frequency isolation → streak validation → beep.

Parameters. The STFT front-end is fixed by the telephony stream. The detection thresholds are what we tune (Table 6). Only four thresholds drive the decision which includes the dominant-peak margin, the energy-isolation test, the streak length, and the frequency floor, and we set them on the golden set below.

ParameterDefaultRole
STFT front-end
sampling rate8 kHztelephony stream rate
window / hop64 / 32 msfreq. resolution
FFT size512spectral detail
Detection thresholds (tuned on golden set)
dominant-peak margin20 dBM1M_1 vs. M2M_2 selectivity
energy isolationonreject broadband or music
streak required4consecutive frames
frequency floor350 Hzreject low hum
min magnitude0.1ignore low-energy noise

Table 6. Beep detector parameters. The STFT front-end is fixed by the 8 kHz telephony stream. The four detection thresholds are tuned on the golden set.

Choosing the operating point. We tune on the internal golden set (293 beep-positive, 74 negative) and validate false alarms on a large beep-free set (~65k clips), since a spurious beep would wrongly trigger AMD on a live call. Two targets drive the choice: precision must be 1.0 (no false alarms on golden), and the false-alarm rate on the beep-free stream must be as low as possible. Figure 2 sweeps the dominant-peak margin. A loose margin (5–10 dB) catches every beep (recall 1.0) but false-alarms on up to 10% of beep-free clips, 15 dB still leaves 1.4%. Adding the energy-isolation test at 20 dB drops the false-alarm rate to 0.33% at precision 1.0 and recall 0.942 (missing 17/293 positives, all very-low-energy or noise-masked beeps), while 25 dB costs more recall for no gain. So 20 dB + energy isolation is the knee, and our default.

Figure 2. Beep operating-point selection: recall (golden set) vs. false-alarm rate on the beep-free set as the dominant-peak margin varies (streak = 4; log x-axis). The chosen 20 dB + energy-isolation config (star) gives the lowest false-alarm rate at near-top recall, looser margins raise false alarms, tighter ones cost recall. Precision is 1.0 for every margin ≥15\ge15 dB.

Why streak = 4. The streak length trades transient rejection against recall and latency. A streak of 3 keeps recall high but stays trigger-happy (false-alarm rate ∼1.0–1.4%); a streak of 5 oversuppresses (golden recall falls toward 0.75 at tight frequency tolerance) and adds latency. A streak of 4 balances both, cutting the false-alarm rate to 0.44% for a modest recall cost at only 4×32 ms≈128 ms4 \times32\ \mathrm{ms} \approx128\ \mathrm{ms} of confirmation delay. The internal sets cannot be released, for compliance and data-privacy reasons.

Voicemail Keyphrases

The keyphrase detector matches the normalized ASR transcript against a configurable phrase list. We derived the English list from recurring greetings in internal answering-machine call transcripts. For the Telephony-AMD experiment, which spans English, Spanish, French, and German, we added translations of these phrases into the other three languages.

English keyphrases

please leave your name forwarded to voicemail record your message at the tone after the tone after the beep can’t come to the phone voice messaging system automated voice message has been forwarded to when you’ve finished recording when you have finished recording hang up or press unavailable right now

Spanish keyphrases

por favor deje su nombre desviado al buzón de voz grabe su mensaje después de la señal después del tono después del bip no puede atender el teléfono sistema de mensajería de voz mensaje de voz automático ha sido desviado a cuando haya terminado de grabar cuelgue o presione no está disponible en este momento

French keyphrases

veuillez laisser votre nom transféré vers la messagerie vocale enregistrez votre message au signal sonore après la tonalité après le bip ne peut pas répondre au téléphone système de messagerie vocale message vocal automatique a été transféré vers lorsque vous avez terminé l’enregistrement raccrochez ou appuyez n’est pas disponible pour le moment

German keyphrases

bitte hinterlassen Sie Ihren Namen zur Mailbox weitergeleitet nehmen Sie Ihre Nachricht auf nach dem Signalton nach dem Ton nach dem Piepton ist gerade nicht erreichbar Sprachnachrichtensystem automatische Sprachnachricht wurde weitergeleitet an wenn Sie mit der Aufnahme fertig sind legen Sie auf oder drücken Sie im Moment nicht erreichbar

Voicemail Detector Prompt

Answering-machine detection prompt (frozen)

The LLM detector runs over the live transcript and raises the AMD pending signal when the far end shows automated-mailbox characteristics. The prompt below is the frozen version used for every voicemail number reported in the paper. It differs from the one shipped in our production pipeline in one respect, argued in § 5.1: the positive class is any automated mailbox system answering on the callee’s behalf, not only one that invites a recording. The detector additionally returns which state it observed, so that the downstream action can separate “retry later” from “stop calling this number” without collapsing the two at detection time. Both prompts are reproduced verbatim except that personal and company names inside the illustrative examples are shown as ⟨name⟩ and ⟨company⟩.

You classify the callee side of an outbound phone call for answering-machine detection (AMD).

Call the answering_machine_detector tool whenever an automated voicemail, answering machine, or carrier mailbox system has answered in place of the person being called. The decisive question is whether a recorded or synthetic system answered on the callee’s behalf, not whether a message can actually be left right now.

Set state to whichever applies:

  1. greeting_invites_message, the recording invites a message: a personal greeting (“you’ve reached ⟨name⟩, leave a message after the tone”), a carrier voicemail greeting naming a network or subscriber, or any instruction to record after a tone, beep or signal.

  2. mailbox_unavailable, an automated message reporting that the mailbox or line cannot take the call, so no recording is possible: the mailbox is full, has not been set up, or the subscriber is temporarily unreachable and the caller is asked to try later.

  3. number_invalid, a carrier intercept reporting that the number itself is unusable: not in service, no longer assigned, or the call cannot be completed as dialled.

These messages occur in any language. Classify by what the message does, not by matching English wording.

Do NOT call the tool when:

  • A live person answers conversationally, greets the caller, asks a question, or engages in two-way dialogue, including brief or abrupt answers

  • The far end is an IVR menu, a queue or hold announcement, or a recorded business greeting that routes the call rather than taking a message. If a recording offers menu options and also mentions leaving a message, treat the menu as dominant and do NOT call the tool

  • The transcript is empty, untranscribed, or too garbled to identify

When the transcript is genuinely ambiguous between a live person and a recording, do NOT call the tool.

IVR Detector Prompt

The IVR detector is a second tool exposed to the conversational LLM, distinguished from the voicemail detector by what it is asked to look for. Where the voicemail prompt keys on a mailbox system answering for the callee, this one keys on an IVR navigation: a menu, queue or recorded greeting is a far end the agent should stay on and traverse, not one it should record against.

IVR detection prompt

You classify the callee side of an outbound phone call for interactive voice response (IVR) detection.

Call the IVR detector tool when the transcript shows any automated, pre-recorded or synthesised system message from the callee’s organisation:

  • A menu offering choices (“press one for . . .”, “say sales or say support”)

  • An instruction to enter digits (“enter your account number”, “followed by the pound key”)

  • A queue or hold announcement (“your call will be answered in the order it was received”, “please stay on the line”, “estimated wait time”)

  • A recorded greeting or business-hours announcement (“thank you for calling ⟨company⟩”, “our offices are currently closed”)

  • A monitoring or recording notice delivered as a recorded system message, or a language-selection prompt

The decisive question is whether a recorded or synthetic system is speaking on behalf of an ORGANISATION, not whether a menu is offered: a greeting or hold message with no selectable option still counts, provided it is an organisation’s system answering its own line.

Do NOT call the tool when:

  • The far end is an assistant answering for an individual rather than an organisation: it offers to take the caller’s name and reason and see whether that person is available, or asks the caller to hold while it fetches them. That is a personal call screener (class 4), not an IVR, even though it is automated and may ask the caller to stay on the line

  • The far end invites the caller to leave a recorded message, that is voicemail

  • A live person answers conversationally, even if they say “thank you for calling”, give a company name, or mention recording

  • The transcript is empty, untranscribed, or too garbled to tell

  • You are uncertain, when in doubt, do NOT call the tool

Tuning. The prompt was iterated on the 119-clip tuning split of § 4.1 only; the 281-clip evaluation split was not inspected during prompt development.

Signal-Layer Panel

The panel has five metrics, each a summary of the far-end signal over the normalised window. Formant-band concentration is the fraction of spectral energy in 200–800 Hz, where the first formants of voiced speech sit. Voicing continuity is the fraction of frames judged voiced by a normalised-autocorrelation pitch test. Level dynamics is the mean frame RMS over 50 ms windows. Background flatness is spectral flatness, the ratio of the geometric to the arithmetic mean of the spectrum, which is high for noise-like audio and low for tonal audio. Segment structure is the mean duration of a contiguous speech segment under an energy-gated VAD.

Each is calibrated on human speech alone, by the rule of § 4, before being compared against synthesis. Table 7 reports two null values per metric. The within column is the 99th percentile of ∣g∣|g| over 200 random splits of a single human corpus, which is the noise floor: how large ∣g∣|g| gets when there is no real difference. The cross column is ∣g∣|g| between the two independent human corpora, which should stay at that floor if the metric measures the speaker rather than the corpus. A metric is admitted only when the cross value does, at ∣g∣≤1|g| \le1. Four fail, each by a wide margin, and all four correlate with realised speech fraction while the survivor does not.

Metricwithin p99crossadm.
Formant-band conc.0.251.24no
Voicing continuity0.252.79no
Level dynamics0.211.11no
Background flatness0.283.95no
Segment structure0.230.26yes
held-out test (n+=257n_{+}=257, n−=1324n_{-}=1324)
ROC-AUC vs. human0.624 [0.597, 0.652]
operating points, threshold fitted on train
FPR budgettest FPRrecallFP
1%1.06%0.00014
2%1.74%0.00023
5%3.25%0.00443
Robocall control, AUC vs. human0.414

Table 7. Signal-layer panel with a held-out threshold (Section 5.3).

The surviving metric is then fitted on a training half and read once on a held-out half split by source call, with the threshold set as described in Appendix § 20. It reaches 0.624 ROC-AUC and has no usable recall at a 1–2% false-positive budget. The robocall control, which is pre-recorded human speech, scores below the human negatives (0.414), so the metric is weakly specific to synthesis and not merely to audio that is not a live call. It is still too weak to act on.

Arbitration Sweep

The priority order is swept on the in-house held-out half, which carries all four classes; the public split cannot exercise it, because it contains no AI callee and so cannot produce the conflicts the ordering exists to resolve. Table 8 reports all six orderings of the three automated classes at a 20 s window (n=501n=501).

Priority orderH F1H FPVM F1AI F1Acc
VM > IVR > AI0.97970.9460.5710.938
VM > AI > IVR0.97970.9460.5710.938
IVR > VM > AI0.97970.9440.5710.936
AI > VM > IVR0.97970.7370.4050.798
AI > IVR > VM0.97970.7330.4050.796
IVR > AI > VM0.97970.7330.4050.796

Table 8. Priority sweep on the in-house held-out half (n=501n=501, 20 s window). H FP counts live humans given an automated label, out of 241. HUMAN is the default throughout and is never outranked. The human decision is invariant; the automated classes turn on whether voicemail outranks AI callee.

Two things hold across the sweep. The human-versus-automated decision does not move at all: F1 0.979, precision 0.987, recall 0.971, and exactly 7 of 241 live humans given an automated label under every ordering. And the ordering matters only through the voicemail-versus-AI-callee relation. With voicemail ranked first, four-way accuracy is 0.936 to 0.938; invert it and accuracy falls to 0.796 to 0.798, as 88 voicemail calls are read as AI callee rather than 18 AI callees read as voicemail. Voicemail recall drops from 0.982 to about 0.59 and AI callee precision from 0.762 to 0.260. The position of IVR is immaterial at 0.2 points, which is what two calls can support.

Table 9 sweeps the decision window and the ordering on the public evaluation half, where the window effect is measurable: it moves accuracy by 3.9 points. The ordering is uninformative there for the reason above, and in any case only three of 281 clips trigger two components, with one of those firing both at the same timestamp to the hundredth of a second, so “first” is decided by implementation order. We report those policies as equivalent on that split and draw the ordering conclusion from the in-house sweep instead.

OrderWWAccPrecH\mathrm{Prec}_{H}RecH\mathrm{Rec}_{H}FPH\mathrm{FP}_{H}
VM →\rightarrow IVR →\rightarrow AI5 s94.30.8481.0000
VM →\rightarrow IVR →\rightarrow AI15 s98.20.9631.0000
VM →\rightarrow IVR →\rightarrow AI30 s98.20.9631.0000
IVR →\rightarrow VM →\rightarrow AI30 s98.60.9631.0000
first-firing (no priority)30 s98.90.9631.0000

Table 9. Arbitration on the public evaluation half. FPH\mathrm{FP}_{H} counts live humans given an automated label. No policy mislabels a human at any window.

Cadence Classifier

The AI callee detector is a logistic-regression classifier over seven timing features taken from the far end’s first turn: the median, standard deviation, coefficient of variation and 90th percentile of the gaps between consecutive words; the articulation rate in words per second; and the median and coefficient of variation of word durations. Word start and end times come from the deployment’s ASR. Features are standardised on the training half, and the weights are fitted there by full-batch gradient descent on the logistic loss (4,000 steps, learning rate 0.08), with AI CALLEE as the positive class and HUMAN as the negative.

The decision threshold is set on the training half at a 5% false-positive budget against the human negatives, and is applied unchanged to the held-out half. A call whose opening far-end turn carries fewer than four timed words receives no score and cannot fire: absence of evidence, which is the correct behaviour for a component that reads timing, and the reason the negatives in Table 2 are the 62 of 481 human far ends whose opening turn is long enough to score. Table 2 reports individual ROC-AUCs for four of the seven features; the fused model uses all seven.

Signal-Layer Operating Point

The threshold for the signal-layer detector is fitted on the training half and applied unchanged to the held-out half. Clips are split 40/60 by source call, stratified by corpus and, for the AppTek arm, by accent group, so that no call contributes to both halves and the accent mix is matched. Training comprises 533 Harper Valley, 348 AppTek, 172 VoiceBench and 225 robocall clips; test comprises 800, 524, 257 and 338 respectively.

For a target human false-positive budget bb, the threshold is the ⌊b⋅N−⌋\lfloor b \cdot N_{-} \rfloor-th highest fused score among training human clips. That threshold is then applied to the test half without adjustment, and Table 7 reports the realised test FPR beside the budget it was set to; the two agree to within a point in every case, which is the check that the training human score distribution transfers. Recall is the fraction of test TTS clips scoring above the same threshold.

We report budgets of 1%, 2% and 5% and deliberately omit tighter ones. With 1,324 human clips at test, a 0.5% budget corresponds to roughly six admissible false positives, and the sampling interval on a rate estimated from six events is wider than the differences between the operating points being compared. A tighter budget can be quoted only with an order of magnitude more human data, which is the practical reason the human arms, not the positive arm, are the ones worth growing.

References

  1. [1]AbijahKaj. 2024. Telephony-amd dataset. Hugging Face Datasets.
  2. [2]Kemal Altwlkany, Sead Delalić, Elmedin Selmanović, Adis Alihodžić, and Ivica Lovrić. 2024. A recurrent neural network approach to the answering machine detection problem. arXiv preprint arXiv:2410.08235.DOI
  3. [3]Nikolay Anisimov, Mikhail Olkhovsky, and Konstantin Kishinsky. 2017. Answering machines detection in contact centers using convolutional neural networks. Technical report, Bright Patterns, Inc., San Bruno, CA, USA. Technical report; no publication date stated.
  4. [4]AppTek. 2024. AppTek multi-accent call center dialogues corpus. Hugging Face Datasets.
  5. [5]Asterisk Project. 2024. Amd — answering machine detection application. Asterisk Documentation.
  6. [6]California State Legislature. 2018. Senate bill 1001, bots: Disclosure (bolstering online transparency act). Cal. Bus. & Prof. Code §§ 17940–17943.
  7. [7]Nuria Alina Chandra, Hannah Lee, Ryan Murtfeldt, Lin Qiu, Arnab Karmakar, Emmanuel Tanumihardja, Kevin Farhat, Ben Caffee, Changyeon Lee, Jong-wook Choi, Sejin Paik, Aerin Kim, and Oren Etzioni. 2025. Deepfake-eval-2024: A multi-modal in-the-wild benchmark of deepfakes circulated in 2024. arXiv preprint arXiv:2503.02857.arxiv.org/abs/2503.02857
  8. [8]Yiming Chen, Xianghu Yue, Chen Zhang, Xiaoxue Gao, Robby T. Tan, and Haizhou Li. 2024. VoiceBench: Benchmarking LLM-based voice assistants. arXiv preprint arXiv:2410.17196. Dataset: https://huggingface.co/datasets/hlt-lab/voicebench.arxiv.org/abs/2410.17196
  9. [9]Cisco. 2017. Answering machine detection algorithm. Cisco Collaboration Knowledge Base.
  10. [10]James W. Cooley and John W. Tukey. 1965. An algorithm for the machine calculation of complex Fourier series. Mathematics of Computation, 19(90):297–301.DOI
  11. [11]European Parliament and Council of the European Union. 2024. Regulation (eu) 2024/1689 laying down harmonised rules on artificial intelligence (artificial intelligence act). Official Journal of the European Union. Article 50, transparency obligations for providers and deployers of certain AI systems.
  12. [12]Jon Ander Gómez, Luis M. Montes-Novella, and Jesús García. 2014. Answering machine detector for call centers. In Proceedings of IberSPEECH 2014, page 198, Las Palmas de Gran Canaria, Spain.
  13. [13]Fredric J. Harris. 1978. On the use of windows for harmonic analysis with the discrete Fourier transform. Proceedings of the IEEE, 66(1):51–83.DOI
  14. [14]Infobip. 2024. Reduce call costs with answering machine detection. Infobip Documentation.
  15. [15]Guan-Ting Lin, Jiachen Lian, Tingle Li, Qirui Wang, Gopala Anumanchipalli, Alexander H. Liu, and Hung-yi Lee. 2025. Full-duplex-bench: A benchmark to evaluate full-duplex spoken dialogue models on turn-taking capabilities. arXiv preprint arXiv:2503.04721.arxiv.org/abs/2503.04721
  16. [16]Dominik Macháček, Raj Dabre, and Ondřej Bojar. 2023. Turning Whisper into real-time transcription system. In Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (IJCNLP-AACL): System Demonstrations, pages 17–24.
  17. [17]Dominik Macháček and Peter Polák. 2025. Simultaneous translation with offline speech and LLM models in CUNI submission to IWSLT 2025. In Proceedings of the 22nd International Conference on Spoken Language Translation (IWSLT 2025).arxiv.org/abs/2506.17077
  18. [18]Enxhi Minaj. 2020. Optimization of answering machine detection application in asterisk. Master’s thesis, Lappeenranta–Lahti University of Technology (LUT).
  19. [19]OpenAI. 2024. GPT-4o system card.
  20. [20]Sathvik Prasad and Bradley Reaves. 2023. Robocall audio from the FTC’s project point of no entry. Technical Report TR-2023-1, North Carolina State University.
  21. [21]Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. Proceedings of the 40th International Conference on Machine Learning (ICML), pages 28492–28518.arxiv.org/abs/2212.04356
  22. [22]Antoine Raux and Maxine Eskénazi. 2008. Optimizing endpointing thresholds using dialogue features in a spoken dialogue system. In Proceedings of the 9th SIGdial Workshop on Discourse and Dialogue, pages 1–10.DOI
  23. [23]Kumar Saurav. 2026. Real-time voicemail detection in telephony audio using temporal speech activity features. arXiv preprint arXiv:2604.09675.arxiv.org/abs/2604.09675
  24. [24]Aastha Sharma and Guangjing Wang. 2026. VoxENES 2026: Benchmarking generalization of speech spoofing detectors against LLM-era TTS and voice conversion. arXiv preprint arXiv:2607.11706.arxiv.org/abs/2607.11706
  25. [25]Sinch. 2024. Answering machine detection (amd v2). Sinch Voice API Reference.
  26. [26]Gabriel Skantze. 2021. Turn-taking in conversational systems and human-robot interaction: A review. Computer Speech & Language, 67:101178.DOI
  27. [27]Telnyx. 2025. Answering machine detection. Telnyx Programmable Voice Documentation.

Paper details

vlnc:W_6xkzxqdv
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC0-1.0
Usage

views · 7

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents