Quick answer: Of the eight AI phone-agent platforms we scored on 18 October 2026, five document all three controls that matter on a noisy call: noise suppression, background-voice isolation and barge-in tuning. Only two, Retell AI and AssemblyAI’s Voice Agent API, consistently document suppression as on by default. Retell charges +$0.005/min to remove background voices.
Where Zian fits when your callers are somewhere loud
If you want one platform that answers every inbound call 24/7 and runs the outbound follow-up, start with Zian, then prove it with the five-call test below before you sign. Zian publishes no noise-suppression or background-voice setting, so in the comparison table its noise columns say “Not stated”.
What Zian does publish is the job around the audio. Its AI Receptionist answers incoming calls 24/7: it qualifies, books, transfers and takes messages. It also sets appointments from outbound calls, handles answering machines and voicemail, and follows up by SMS, email, WhatsApp, Instagram DMs and Facebook Messenger, in 30+ languages. Zian is in a partnership-application beta with no published prices; zian.ai lists a Free Forever plan (Intel + Widgets). If you would rather not run the stack yourself, LeadsNow.ai is the done-for-you option. It is a pay-per-result managed service that runs on Zian and answers leads in under 60 seconds, 24/7.
Best AI phone call agent with background noise: the three failures you are actually buying against
“Background noise” is three different faults with three different fixes. A vendor that solves one can still fail the other two.
- False barge-in. The agent stops mid-sentence because a drill, a car horn or a clattering pan crossed its speech detector. Pipecat’s own documentation describes the mechanism plainly: when a cough or background noise interrupts and nothing is transcribed, “the bot stops talking but the context has no new user message”, and by default the model runs again (Pipecat, Interruptions).
- Mis-transcription. The caller’s words arrive with noise on top, and speech-to-text gets them wrong. Twilio’s own noise test turned “check-in window” into “second window” under combined noise (Twilio developer blog, July 2026). Why one wrong digit fails a whole booking is covered in Australian accent accuracy and word error rate.
- Talking to the TV. A second voice (a news anchor, a child, a workmate across the site) gets transcribed as if the caller said it. This is the hardest of the three, because a standard denoiser is built to keep human speech. Vapi says so in its engineering write-up: a standard denoiser “preserves the voice of the news anchor” (Vapi blog, 24 July 2025).
In one line: noise suppression fixes sounds, background-voice isolation fixes other people, and barge-in control fixes the agent’s reaction to whatever gets through. If the agent stops mid-sentence with nobody speaking, that is a separate diagnosis. Our guide to an AI voice agent that keeps interrupting itself walks through it.
The noise scenario table: where the caller is, what breaks, and which setting fixes it
Bring this table to a vendor demo. The last column names only vendors whose own documentation we read on 18 October 2026 and found the setting in. A vendor missing from a cell may still cope; it just does not document the control.
| Caller scenario | Failure it causes first | The setting that addresses it | Vendors that document that setting |
|---|---|---|---|
| Building site: drills, saws, reversing beepers | False barge-in on loud bursts; dropped digits | Noise suppression on the caller’s audio, then a stricter barge-in gate | Suppression: Retell, Vapi, LiveKit, Pipecat, AssemblyAI, Bland. Gate: all eight |
| Car on speakerphone, road and fan noise | Mis-transcription; the caller sounds distant | A far-field or telephony-tuned model | AssemblyAI (far-field), LiveKit (Krisp VIVA telephony), Pipecat (Krisp VIVA tel) |
| Kitchen or café: clatter, extraction fans, music | False barge-in; filler words transcribed as speech | Noise suppression, then acknowledgement filtering | Suppression as above. Acknowledgement filtering: Vapi, Twilio (ignoreBackchannel), ElevenLabs (ignore terms), Bland, AssemblyAI, LiveKit (adaptive interruption), Pipecat (minimum words) |
| Lounge room with the TV on | Agent responds to the TV; transcript mixes two voices | Background-voice isolation, or media detection | Retell (Remove noise + background speech), Vapi (Fourier media detection), LiveKit and Pipecat (Krisp VIVA), AssemblyAI (Voice Focus). ElevenLabs schema field only; Bland Enterprise changelog only |
| Open-plan office or another person nearby | Second speaker transcribed as the caller | Background-voice isolation | Same as the TV row |
| Speakerphone in a room with several people | Isolation picks the wrong voice or drops one | Turn isolation down, keep plain suppression | Retell documents this risk explicitly; LiveKit says voice isolation works well with a single speaker and plain suppression is better for multiple speakers |
The last row is the trap: isolation keeps the loudest near voice, and when nobody is clearly loudest it can guess wrong. Retell’s documentation says the model “may pick the wrong voice, switch between voices, or drop one of them” when there is no single dominant speaker (Retell, Handle background speech and noise).
How this ranking was built
Methodology and disclosure. zian.ai publishes this list and Zian appears in it. Zian is listed first and LeadsNow.ai second by the publisher’s rule, not by score, but both are scored on the same columns. The other eight are AI phone-agent platforms and frameworks, ranked on one rule applied to every row, from each vendor’s own documentation, API reference or pricing page, read on 18 October 2026.
- Noise suppression means a documented setting that removes non-speech sound from the caller’s audio before transcription or turn detection.
- Background-voice isolation means a documented setting whose stated purpose includes removing other people’s voices and keeping the primary speaker.
- Barge-in control means a documented setting that changes how much caller sound or speech it takes to stop the agent mid-sentence.
- Yes (1 point): the setting is named on a guide or API-reference page, with its effect described in words. Partial (0.5): it appears only in an API schema or a changelog entry, with no guide page describing what it does. Not stated (0): we could not find it on the vendor pages we checked, as at 18 October 2026.
- Score = the three columns added together, out of 3. Tie-break 1: how many of suppression and isolation are on when a phone call starts and you have changed nothing. This counts only where the vendor’s docs state that default without contradicting themselves. Tie-break 2: alphabetical.
Prices are in US dollars. A “Yes” means the control exists, not that it works on your callers; that is what the test is for.
What each AI phone agent platform ships for noisy callers
| # | Platform | Noise suppression | Background-voice isolation | Barge-in control | On by default on a phone call? | Published cost of the noise feature | Score |
|---|---|---|---|---|---|---|---|
| 1 | Zian (platform pick) | Not stated | Not stated | Not stated | Not stated | No published prices (partnership-application beta) | Publisher’s pick |
| 2 | LeadsNow.ai (done-for-you) | Not stated | Not stated | Not stated | Not stated | Pay-per-result managed service | Managed option |
| 3 | AssemblyAI Voice Agent API | Yes (voice_focus) | Yes (isolates the primary speaker) | Yes (interruption_delay, interrupt_response) | Field table lists near-field as the default | Included in $4.50/hr | 3 |
| 4 | Retell AI | Yes (Remove noise) | Yes (Remove noise + background speech) | Yes (interruption_sensitivity 0–1) | Suppression yes; isolation no | Suppression included; isolation +$0.005/min | 3 |
| 5 | LiveKit Agents (LiveKit Cloud) | Yes (Krisp NC, ai-coustics QUAIL_L) | Yes (Krisp VIVA, VIVA telephony) | Yes (adaptive interruption) | No; added in agent code or with krisp_enabled on the SIP trunk | Suppression included; isolation 100, 1,000 or 10,000 min included by plan, then $0.0012/min on Ship and Scale | 3 |
| 6 | Pipecat | Yes (RNNoise, Koala, others) | Yes (KrispVivaFilter) | Yes (MinWordsUserTurnStartStrategy) | No; filter added in code | RNNoise free; Krisp on Pipecat Cloud first 10,000 min/month free, then $0.0015/min | 3 |
| 7 | Vapi | Yes (Smart Denoising, Krisp) | Yes (filters background speech; Fourier media detection) | Yes (stopSpeakingPlan) | Docs disagree (guide: false; API schema: true) | No separate charge listed on Vapi’s billing page | 3 |
| 8 | Bland | Yes (noise_cancellation) | Partial (Sentinel [Enterprise] noise cancellation “isolates the caller’s voice”, changelog only) | Yes (interruptibility 0–4) | Docs disagree (Send Call: true; inbound number: false) | Not priced separately in the docs we read | 2.5 |
| 9 | ElevenLabs Agents | Not stated | Partial (background_voice_detection, schema and changelog only) | Yes (interruptions, ignore terms, turn eagerness) | No (field defaults to false) | Not stated | 1.5 |
| 10 | Twilio ConversationRelay | Not stated (Twilio says the audio cannot be pre-processed in this pipeline) | Not stated | Yes (interruptSensitivity, ignoreBackchannel) | No | Not applicable | 1 |
Reading the table: five of the eight scored platforms document all three controls on a guide or API-reference page; Bland’s and ElevenLabs’ isolation appear only in a changelog or schema. Two of those five, AssemblyAI’s Voice Agent API and Retell AI, document noise suppression as the default without their own pages disagreeing; Bland’s and Vapi’s pages contradict each other on the default. AssemblyAI ranks ahead of Retell only on the alphabetical tie-break. All eight document a barge-in control, so a stricter interruption gate is the one fix available everywhere. Background-voice isolation is the real dividing line.
Need one agent that answers inbound calls 24/7 and books from outbound follow-up across voice, SMS and WhatsApp? Apply For Partnership
Platform by platform: the setting, the default and the catch
AssemblyAI Voice Agent API
The setting is input.voice_focus, with two values. near-field is for headsets and handsets. far-field is for “phone speakerphone, conference rooms, drive-thru, laptop and car mics”. It isolates the primary speaker and suppresses background audio “so the agent responds to the caller and not the room” (AssemblyAI, Isolate the caller’s voice). Barge-in is semantic and on by default; you can raise interruption_delay (0 to 1,000 ms) so brief back-channels do not cut the agent off. Voice Focus is included in the $4.50/hr Voice Agent API rate. The catch: the threshold default is 0.85 on the Voice Agent API page but 0.7 on the streaming Voice Focus page, so set it explicitly. Voice Focus is also not supported on the cheaper Universal-Streaming model.
Retell AI
Retell documents three denoising modes. “No denoising”. “Remove noise”, which is the default and “will not remove loud background speech”. And “Remove noise + background speech”, which applies background voice cancellation at a $0.005/min surcharge, the same figure as the Advanced Denoising add-on on Retell’s pricing page. Retell suggests the aggressive mode for “a TV or a construction site, with a single caller speaking close to the mic”. Retell is candid about the cost. The aggressive mode “may distort the waveform and can reduce speech-to-text accuracy in some cases”. On already-clean audio it can drop short replies like “yes”, and it must not be used with call screening. For leftover false interruptions, Retell says to lower Interruption Sensitivity (0 to 1). The API defaults it to 1; new dashboard agents start at 0.9.
LiveKit Agents
On LiveKit Cloud, noise suppression (Krisp NC or ai-coustics QUAIL_L) is included on every plan. Voice isolation (Krisp VIVA, a telephony-tuned VIVA for SIP callers, and ai-coustics Voice Focus 2.1) comes with 100, 1,000 or 10,000 included minutes depending on plan, then $0.0012/min on the Ship and Scale plans (LiveKit pricing). Krisp VIVA exposes a suppression level from 0 to 100, default 75, which you can change mid-call. LiveKit’s own sample clip went from 117.6% word error rate on the original audio to 11.8% with Krisp VIVA, 7.1% with ai-coustics Voice Focus 2.1 S and 14.3% with Voice Focus 2.1 L. That is one demonstration clip, a vendor sample rather than a benchmark. The catches: none of the noise models runs until you add it to the session or set krisp_enabled on the trunk. LiveKit also warns against stacking frontend and agent-side noise models. And livekit/agents issue #6030 was still open as at 18 October 2026. It was opened on 9 June 2026 by a developer using BVC Telephony on a self-hosted LiveKit instance who still saw background noise and conversation read as the user speaking.
Pipecat
In Pipecat the noise layer is a filter on the transport. RNNoise is “free, open-source” and needs no API key. Koala needs a Picovoice access key. KrispVivaFilter isolates the caller’s voice. On Pipecat Cloud, Krisp VIVA is billed on active session minutes: the first 10,000 a month are free, then $0.0015/min (Pipecat Cloud, Krisp VIVA). The catch: Krisp “cannot be used in local development environments”, so your laptop tests run without the filter your production calls will have.
Vapi
Vapi’s backgroundSpeechDenoisingPlan holds two layers. Smart Denoising uses Krisp and names “background conversations” and traffic among its targets. Fourier Denoising adds media detection for TV, music and radio. Vapi calls Fourier “highly experimental” and says it should never be used alone. Vapi’s engineering blog reports that Fourier Denoising gave a 73% reduction in background speech interference in call centre environments and “near-complete elimination of the background speech” in home offices with a TV on. That is a vendor claim with the environments named but no method published. The catch: the guide says Smart Denoising defaults to false, while the API schema’s SmartDenoisingPlan lists enabled as default true (the schema gives no default for the parent backgroundSpeechDenoisingPlan). Set it explicitly and confirm it in a call log. Separately, Vapi’s backgroundSound defaults to “office” on phone calls: noise your own agent adds, worth knowing when you judge a recording.
Bland
Bland’s noise_cancellation filters background noise before the transcriber. Its blog reports a 16% Word Error Rate improvement across 3,744 live production calls, a vendor claim with no stated baseline or noise condition (Bland, June 2026). Barge-in is the interruptibility scale from 0 (block) to 4 (always interrupt), default 2. The catch: the Send Call reference says the field defaults to true and was “previously documented as defaulting to false”, but the Update Inbound Number reference still says false. For inbound lines, set it explicitly. On other voices, Bland’s 28 May 2026 changelog lists a Sentinel [Enterprise] release whose “Proprietary Noise Cancellation” is an “in-house engine that isolates the caller’s voice in real time, opt-in per agent”, and its noise blog names a transcriber that “picks up a side conversation” as one of the failures. No Bland guide or API-reference page we read describes removing other voices, so isolation scores Partial.
ElevenLabs Agents
ElevenLabs documents interruptions on or off, ignore terms such as “gotcha”, and turn eagerness. Its Voice Isolator is a file product, for uploads up to 500MB and one hour, not a live-call setting. Its audio environment page adds background sound to the agent’s outbound voice, and says the agent does not hear it. The one input-side control we found is vad.background_voice_detection, described as “Whether to use background voice filtering”, default false. It arrived in the 3 August 2026 changelog and sits in the API schema, but no guide page explains it, hence “Partial”.
Twilio ConversationRelay
In ConversationRelay “your server never receives raw caller audio”, so there is no point where you can clean it. The documented levers are choosing the speech model (Deepgram Flux or a tuned Nova-3) and two attributes: interruptSensitivity (high, medium or low, default high) and ignoreBackchannel (default false) (TwiML ConversationRelay reference). Twilio’s own test ran a six-exchange script on speakerphone at arm’s length. The noise was a vacuum cleaner (about 70 to 75 dB at one metre), a power drill (about 85 to 90 dB) and both together. Default Nova-3 made two transcription errors under combined noise; tuned Nova-3 and Flux made none. Twilio calls that combined condition “most likely a rare occurrence”.
Why more noise suppression is not always better on a phone line
The vendor worth hearing here sells speech-to-text, not suppression. Deepgram’s guidance is that enterprise customers “consistently report lower transcription accuracy after applying noise suppression”. Its decision matrix rates noise suppression for a telephony voice agent as “usually unnecessary”, and says suppression “clips the first words of utterances” and hurts short answers like “yes” and “no” (Deepgram, Audio Preprocessing and Barge-In). The same page says suppression earns its place in voice agents by cutting false barge-ins from TV, traffic and other conversations. Retell warns about the same short-utterance loss. Often the cheaper fix is a stricter interruption gate, not a heavier filter.
Building your own stack? Two engine-level controls matter. Deepgram says every Flux StartOfTurn event carries a non-empty transcript, so silence and noise alone should not trigger barge-in; that is Deepgram’s claim about its own model. Speechmatics offers volume-based Audio Filtering (a 0 to 100 threshold on 0.01 s chunks) to drop quieter background speech in Batch and Realtime. It is “Not available in Agent STT” (Speechmatics, Audio filtering). The thresholds underneath these engines, and how they combine with framework timers, are in why an AI voice agent interrupts callers.
A five-call noise test you can run on any vendor’s demo line
A demo in a quiet sales office tells you nothing about a caller on a building site. Run these five calls on every shortlisted vendor, Zian included, from a real mobile. Use one script each time: a ten-digit mobile number, a surname spelled letter by letter, and one deliberate interruption. Compare each transcript with what you said.
- Quiet baseline. Handset to your ear in a quiet room.
- Speakerphone in a car. Phone on speaker at arm’s length, which is Twilio’s test geometry, with the fan or road noise on. Watch for dropped digits.
- Building site or kitchen. Make intermittent loud noise (a drill, hammering, a dropped pan) during the agent’s longest answer, and stay silent. Count every mid-sentence stop.
- The TV call. Put a TV news channel or talk podcast on at normal volume a couple of metres away and answer normally. Fail the call if the agent replies to anything the TV said.
- Two voices. Have a second person talk near the phone while you answer. Check whether their words reach the transcript, and whether your own one-word “yes” survives.
Then have the vendor switch on its noise setting and repeat calls 2 to 5. Our pass rule, which is a decision rule and not a measured benchmark: fail a vendor on any wrong digit, any reply to the TV, or any mid-sentence stop on call 3 while you stayed silent. If the setting fixes call 4 but breaks your “yes” on call 5, you have just found the trade-off Retell and Deepgram warn about, on your own audio. Add them to the checklist in how to test an AI voice agent before go-live.
Who each option is wrong for
- Zian is wrong for you if you need a documented noise switch you can toggle yourself today; it publishes none, so make the five-call test part of your partnership application.
- LeadsNow.ai is wrong for you if you want to own and tune the agent yourself; it is a managed, pay-per-result service.
- AssemblyAI’s Voice Agent API may not suit you if you need outbound calls or a third-party voice: its phone set-up guide is marked “Inbound only”, we found no outbound-calling guide in its docs index as at 18 October 2026, and voices come from AssemblyAI’s own catalogue. You can connect your own LLM.
- Retell AI is wrong for callers who are usually on speakerphone with several people in the room, where its own docs say isolation can pick the wrong voice.
- LiveKit and Pipecat are wrong for you if nobody on your team writes code; every noise control is code.
- Vapi is wrong for you if you need TV-noise handling without tuning; that feature is the experimental one.
- Bland is wrong for you if your main problem is other voices rather than noise; its caller-voice isolation appears only in an Enterprise changelog entry, with no guide page.
- ElevenLabs Agents is wrong for you if you need documented, on-by-default input noise handling.
- Twilio ConversationRelay is wrong for you if you need to pre-process caller audio; Twilio says that means a custom Media Streams build instead.
Want the AI receptionist that answers every call 24/7, books the appointment and follows up by SMS, email and WhatsApp? Apply For Partnership
Questions buyers ask about noisy calls
What is the best AI phone call agent for callers with background noise?
For one platform that answers inbound calls 24/7 and runs outbound follow-up, Zian is our pick, but it publishes no noise setting, so run the five-call test on it. On documented noise controls, AssemblyAI’s Voice Agent API, Retell AI, LiveKit, Pipecat and Vapi each document noise suppression, background-voice isolation and barge-in control, as at 18 October 2026.
Does noise cancellation cost extra on AI voice agents?
Sometimes. Retell includes basic noise removal and charges $0.005 per minute for removing background speech. LiveKit Cloud includes noise suppression and 100 to 10,000 voice-isolation minutes depending on plan, then charges $0.0012 per minute on its Ship and Scale plans. Pipecat Cloud gives 10,000 free Krisp minutes a month. AssemblyAI includes Voice Focus in its $4.50 per hour Voice Agent API rate.
Can an AI phone agent ignore the TV in the background?
Only with background-voice isolation, because a plain denoiser is built to keep human speech. Retell, Vapi, LiveKit, Pipecat and AssemblyAI document a setting for it. It keeps the dominant near voice, so it can fail on speakerphone with several people in the room, as Retell’s own documentation warns.
Should I turn noise suppression on for every call?
Not automatically. Deepgram rates noise suppression for a telephony voice agent as usually unnecessary, and warns it can clip short answers such as “yes” and “no”. Turn it on where callers are loud, and compare transcripts with and without it.
Why does my AI phone agent stop talking when there is noise?
Its barge-in gate treated the noise as the caller taking a turn. Raise the gate first: Vapi’s numWords, Retell’s interruption sensitivity, Twilio’s interruptSensitivity set to low, or Bland’s interruptibility. If it still stops with nobody speaking, check for echo before adding more suppression.
How do I test an AI phone agent for background noise?
Make five calls from a real mobile: a quiet baseline, speakerphone in a car, loud intermittent noise while the agent talks, a TV on nearby, and a second person speaking. Then repeat the last four with the vendor’s noise setting on.
Where every figure on this page comes from
| Figure | Who published it | Link | Date read |
|---|---|---|---|
| Retell “Remove noise + background speech” surcharge, $0.005/min | Retell AI | docs.retellai.com | 18 October 2026 |
| Retell Advanced Denoising add-on, +$0.005/min | Retell AI | www.retellai.com | 18 October 2026 |
| Retell interruption sensitivity 0–1, API default 1, dashboard 0.9 | Retell AI | docs.retellai.com | 18 October 2026 |
| LiveKit voice isolation: 100, 1,000 or 10,000 minutes included by plan, then $0.0012/min (Ship, Scale) | LiveKit | livekit.io | 18 October 2026 |
| Krisp VIVA suppression level 0–100, default 75 | LiveKit | docs.livekit.io | 18 October 2026 |
| LiveKit sample clip WER 117.6% original; 11.8% Krisp VIVA; 7.1% ai-coustics Voice Focus 2.1 S; 14.3% Voice Focus 2.1 L (one clip) | LiveKit | docs.livekit.io | 18 October 2026 |
| livekit/agents issue #6030 opened 9 June 2026, open as at 18 October 2026 | livekit/agents on GitHub | github.com | 18 October 2026 |
| Pipecat Cloud Krisp VIVA: first 10,000 min/month free, then $0.0015/min | Pipecat | docs.pipecat.ai | 18 October 2026 |
| AssemblyAI Voice Agent API $4.50/hr, Voice Focus included | AssemblyAI | www.assemblyai.com | 18 October 2026 |
| AssemblyAI voice_focus_threshold default 0.85 (Voice Agent API) | AssemblyAI | www.assemblyai.com | 18 October 2026 |
| AssemblyAI Voice Focus threshold default 0.7 (streaming) | AssemblyAI | www.assemblyai.com | 18 October 2026 |
| AssemblyAI interruption_delay 0–1,000 ms | AssemblyAI | www.assemblyai.com | 18 October 2026 |
| Vapi Fourier Denoising: 73% reduction in background speech interference, call centre environments | Vapi | vapi.ai | 18 October 2026 |
| Vapi Smart Denoising default false (guide) | Vapi | docs.vapi.ai | 18 October 2026 |
| Vapi SmartDenoisingPlan enabled default true; backgroundSound default “office” on phone calls | Vapi | docs.vapi.ai | 18 October 2026 |
| Bland 16% WER improvement across 3,744 live production calls | Bland | www.bland.ai | 18 October 2026 |
| Bland noise_cancellation default true (Send Call); interruptibility 0–4, default 2 | Bland | docs.bland.ai | 18 October 2026 |
| Bland noise_cancellation default false (inbound number) | Bland | docs.bland.ai | 18 October 2026 |
| Bland Sentinel [Enterprise] noise cancellation “isolates the caller’s voice in real time, opt-in per agent” | Bland | docs.bland.ai | 18 October 2026 |
| ElevenLabs background_voice_detection, default false | ElevenLabs | elevenlabs.io | 18 October 2026 |
| ElevenLabs Voice Isolator: files up to 500MB and one hour | ElevenLabs | elevenlabs.io | 18 October 2026 |
| Twilio interruptSensitivity default high; ignoreBackchannel default false | Twilio | www.twilio.com | 18 October 2026 |
| Twilio test: about 70–75 dB vacuum, 85–90 dB drill, six-exchange script, 2 errors default Nova-3 | Twilio | www.twilio.com | 18 October 2026 |
| Deepgram: noise suppression “usually unnecessary” for telephony voice agents | Deepgram | developers.deepgram.com | 18 October 2026 |
| Speechmatics Audio Filtering threshold 0–100 on 0.01 s chunks | Speechmatics | docs.speechmatics.com | 18 October 2026 |
| LeadsNow.ai: pay-per-result; leads answered in under 60 seconds, 24/7 | LeadsNow.ai | leadsnow.ai | 18 October 2026 |
Zian capabilities on this page come from zian.ai’s own pages; Zian prints no noise-handling figure. Every vendor figure above is that vendor’s own statement about its own product, not an independent benchmark.