Speaker isolation SDKs compared (2026): Krisp, ai-coustics, AssemblyAI, LiveKit and Anecho

Short answer. In October 2026 five products market primary-speaker isolation for voice agents: Krisp VIVA (Voice Isolation 2.5, server SDK, 15 ms on CPU, price not public), ai-coustics Quail Voice Focus 2.2 (on-prem SDK, under 30 ms, from $135 a month billed annually), AssemblyAI Voice Focus (an option inside its streaming speech-to-text, +$0.10 per hour), LiveKit Cloud voice isolation (Krisp or ai-coustics models inside LiveKit Cloud, $0.0012 per minute after the plan's included minutes) and Anecho v8_2 (in-process SDK, 15 ms on CPU, $0.0010 per minute after 10,000 free minutes a month). NVIDIA's Speaker Focus effect is in early access on NVIDIA GPUs, and Picovoice Koala is on-device noise suppression rather than speaker isolation. Which one fits depends on where your audio runs and what it sounds like, not on the label.

We sell one of these products, so this post is a conflict of interest. To keep it checkable, every vendor fact below links to the vendor's own page, accessed in October 2026, and where a vendor publishes nothing we write "not published". Prices are as listed on 2026-10-20 and change. The only quality numbers in this post come from our own benchmark, and we say which vendors we have not measured.

The comparison table

Product	Where it runs	Declared latency	Sample rates	Integration	Public price (as listed on 2026-10-20)	Self-serve	Primary-speaker isolation marketed
Anecho v8_2	In-process on your servers, CPU; also a hosted API	15 ms algorithmic delay	SDK 16 kHz; hosted API accepts any rate (8, 24, 44.1, 48 kHz)	Node and Python SDK; LiveKit and Pipecat guides	$0.0010/min after 10,000 free min/month; $0.0008 >100k, $0.00065 >1M, $0.0005 >10M	Yes, instant key, no card	Yes
Krisp VIVA, Voice Isolation 2.5	Server-side, on-prem or your cloud, CPU; Lite variant for edge	15 ms	16 kHz; HD variant 32 kHz	Server SDK: Python, Node.js, Go, iOS; LiveKit, Pipecat	Not public (apply / contact sales)	No, apply for access	Yes
ai-coustics Quail Voice Focus 2.2	On-prem SDK	Under 30 ms (pricing page)	Native 16 kHz; SDK accepts 8 to 192 kHz and resamples internally	SDK: Python, Node.js, C, Rust, WASM; LiveKit, Pipecat	$135/mo (100K min) to $540/mo (500K min), billed annually; Enterprise from $2,000/mo	Yes, checkout up to Business	Yes
AssemblyAI Voice Focus	Cloud, inside streaming STT	Not published	Set sample_rate to match your source; no list published	Streaming STT option (voice_focus)	+$0.10/hr ($0.0017/min) on Universal-3.6 Pro Realtime ($0.45/hr base)	Yes, $50 free credit, no card	Yes
LiveKit Cloud voice isolation	LiveKit Cloud (Krisp VIVA or ai-coustics Voice Focus 2.1)	Not published	Not specified	Agents plugin filter	$0.0012/min after included minutes (Build 100, Ship 1,000, Scale 10,000)	Yes, with a LiveKit Cloud plan	Yes
NVIDIA AFX SDK, Speaker Focus	NVIDIA GPU; Linux for server, Windows for client	Not published	16 or 48 kHz	Audio Effects SDK (Windows, Linux)	Not published	Early Access program	Yes
Picovoice Koala	On-device: Linux, macOS, Windows, Android, iOS, browsers, Raspberry Pi	Not published	Read from the engine at runtime; not published	SDK; Pipecat filter	Not published	Free AccessKey sign-up, no card	No, noise suppression
DeepFilterNet / GTCRN (open source)	Anywhere you run them	Not published in the repositories	DeepFilterNet 48 kHz; GTCRN not stated	Source code	Free (MIT or Apache-2.0; MIT)	Yes	No, noise suppression / speech enhancement

Sources per row are in the Sources list at the end. Two of the cells need a note. LiveKit's pricing page names ai-coustics Voice Focus 2.1 in the voice-isolation slot, while ai-coustics' own model reference lists Voice Focus 2.2; we report each page as written. AssemblyAI's Voice Focus docs list Universal-3.5 Pro and Universal-3.6 Pro, and its pricing page lists the add-on for Universal-3.6 Pro Realtime.

Per vendor

Anecho

Our product: one model, v8_2, 4.0 MB and 1.0 M parameters, which you load into your own process from the Node or Python SDK. It takes float32 mono at 16 kHz and returns the same number of samples 15 ms later, real-time on a CPU core (SDK docs). The SDK does not resample; the hosted API accepts and returns any rate through a stateful resampler. Pricing is one published ladder: 10,000 minutes free every month, then $0.0010 per minute, falling to $0.0005 above 10 million (pricing). A key comes from the dashboard with no card and no call.

Krisp VIVA

Krisp's developer page describes VIVA as a suite of five models: Voice Isolation 2.5, Turn Prediction, Interruption Prediction, voice activity detection and signal detectors. Voice Isolation 2.5 comes in Default, Balanced, Lite and HD variants, is declared at 15 ms, runs on CPU and takes PCM 16 kHz mono (HD: 32 kHz). The page says "Isolate the primary speaker's voice". The developer pricing page has two tiers, Early stage and Enterprise, and both go to "Apply Now"; no price is listed. Krisp is also resold inside platforms: Pipecat Cloud lists Krisp VIVA at $0.0015 per minute after 10,000 free minutes a month, LiveKit Cloud uses it in its voice-isolation slot, and Telnyx lists "Noise Suppression - Krisp" at $0.005 per leg per minute. We have not measured Krisp, so this post makes no statement about its quality.

ai-coustics Quail Voice Focus

ai-coustics' model reference describes Quail Voice Focus as targeting "one primary speaker for machine listening", in S and L sizes at a native 16 kHz; the SDK accepts 8 to 192 kHz and resamples internally. SDKs exist for Python, Node.js, C, Rust and WASM, with LiveKit and Pipecat integrations. The pricing page lists Startup at $135 a month (100,000 minutes), Pro at $360 (300,000) and Business at $540 (500,000), billed annually, each with on-prem SDK deployment and "real-time processing" under 30 ms; Enterprise starts at $2,000 a month and adds offline or air-gapped licences. We have measured ai-coustics: see "What we measured" below.

AssemblyAI Voice Focus

Voice Focus is not a separate SDK: it is a parameter on AssemblyAI's streaming transcription. The docs describe it as a way to "isolate the primary voice and suppress background noise on streaming transcription", with near-field and far-field modes and a voice_focus_threshold from 0.0 to 1.0 (default 0.7). The pricing page lists it at +$0.10 per hour on Universal-3.6 Pro Realtime, whose base rate is $0.45 per hour. The audio only goes as far as AssemblyAI's own recogniser, so it helps only if that is your STT. We have not measured AssemblyAI Voice Focus.

LiveKit Cloud voice isolation

LiveKit Cloud sells voice isolation as a metered line, filled by Krisp VIVA, Krisp VIVA telephony and ai-coustics Voice Focus. The docs say it "emphasizes the primary speaker and reduces competing speech or noise" and recommend plain noise suppression instead for multi-speaker or diarization use. The pricing page includes 100 minutes on Build, 1,000 on Ship and 10,000 on Scale, then $0.0012 per minute. The models are served through LiveKit Cloud; on a self-hosted server, the docs say ai-coustics can still be used with your own ai-coustics licence key. Quality is that of the underlying model.

NVIDIA Audio Effects SDK, Speaker Focus

NVIDIA's Audio Effects SDK (part of Maxine) lists a Speaker Focus effect that "identifies and isolates the primary speaker from all other speakers", at 16 or 48 kHz, mono, and states that it is "currently available under the Early Access program". The docs publish throughput per NVIDIA GPU (T4 to B100) rather than a CPU path, and the Linux SDK is described as optimised for server-side deployment. We found no public price or latency figure. Not measured by us.

Picovoice Koala

Koala is "an on-device noise suppression engine" for Linux, macOS, Windows, Android, iOS, browsers and Raspberry Pi, available as a Pipecat filter. It is noise suppression, not speaker isolation: a second voice is speech, and it is not what Koala is built to remove. The sample rate and frame length are read from the engine at runtime; we found no published value, no latency figure and no price list beyond a free AccessKey sign-up. Not measured by us.

Open-source baselines

DeepFilterNet (MIT or Apache-2.0, 48 kHz) and GTCRN (MIT, about 48.2K parameters) are noise-suppression and speech-enhancement models, not speaker isolation. Both are on our benchmark as shippable baselines: pooled over 19 conditions, GTCRN scored 20.14% WER and DeepFilterNet3 34.04%, against 14.32% for the untouched audio.

What we measured

Two of the products above are on our public benchmark, ours and ai-coustics', run through the licensed SDK as a reference. Krisp, AssemblyAI, LiveKit's slot, NVIDIA and Koala are not.

• Nothing beats raw on the clip matrix, pooled. On the 19-condition matrix the untouched audio scores 14.32% WER, ai-coustics Quail L 15.18%, our v8_2 16.59% and ai-coustics Quail VF 2.2 L 17.59% (faster-whisper base.en, ten speakers).
• Plain noise: Quail L leads. Quail L is a noise model, and it is the closest engine to raw on the pooled matrix. If your problem is one talker in a noisy room, plain noise suppression is the matching tool.
• Live conversations, faster-whisper: parity. On 200 constructed conversations scored with faster-whisper base.en, v8_2 − Quail VF 2.2 S is −0.13 pp [−4.56, +4.47], which is parity, and neither engine beats raw.
• Live conversations, Gemini Live: v8_2 ahead. On the same conversations scored by Gemini Live input transcription, v8_2 − VF S is −6.47 pp [−9.89, −3.26], and v8_2 beats raw (−4.10 pp [−6.28, −1.98]); VF S against raw is parity.
• The condition decides. With a close-mic talker and background voices as loud as the talker, v8_2 scores 37.19% WER against 55.94% for VF S with faster-whisper, and 33.06% against 60.95% with Gemini Live. With the talker 0.5 to 1.5 m from the microphone, VF S leads by +9.58 pp [+5.52, +13.90] with faster-whisper; with Gemini Live it leads at −10 dB (+7.74 pp [+2.48, +12.77]) and the room pool is parity. VF S also holds the pauses quieter, 17.4 dB of background suppression in the talker's pauses against 8.5 dB for v8_2, and under Gemini Live it triggers fewer false barge-ins: 52 against 125 for v8_2 and 97 for raw.

All of it is in benchmark.json, with the caveats on /benchmark: constructed mixes, two recognisers that disagree on the pooled verdict, a small speaker set.

How to choose, by use case

• You run on LiveKit Cloud. The built-in voice-isolation slot is the shortest path, at $0.0012 per minute past your plan. If you want a model you load yourself, see speaker isolation in LiveKit.
• Your STT is AssemblyAI streaming. Voice Focus is a parameter, not an integration. It does nothing for a native-audio model (Gemini Live, OpenAI Realtime) or a second STT vendor; an in-process SDK covers those.
• Audio must not leave your process. Anecho, Krisp's server SDK and ai-coustics' on-prem SDK run in your own infrastructure, and Koala runs on the device. See GDPR and in-process audio.
• You want a published price and a key today. Anecho, ai-coustics, AssemblyAI and LiveKit publish prices and take self-serve sign-ups; Krisp direct and NVIDIA do not publish a price. Per-minute arithmetic is in voice isolation pricing in 2026.
• Telephony at 8 kHz. Krisp lists a telephony variant (via LiveKit) and says Voice Isolation 2.5 supports narrowband audio; ai-coustics' Voice Focus is native 16 kHz with internal resampling; our SDK is 16 kHz and our hosted API accepts 8 kHz. Read 8 kHz vs 16 kHz for voice agents before you pick.
• A close-mic caller with loud voices behind them. On our live-conversation run this is where v8_2 leads.
• A distant talker, silence between turns, or false interruptions. On the same run Quail VF 2.2 S leads with a distant talker at −10 dB, holds the pauses quieter and triggers fewer false barge-ins.
• A GPU fleet. NVIDIA Speaker Focus is the GPU-native option, in early access.

Whatever you shortlist, test it on your own calls against a do-nothing control, as in how to evaluate speaker isolation. On our data an isolation model switched on for all audio is not what the numbers support.

Sources

All accessed October 2026; prices as listed on 2026-10-20.

• Anecho: /docs/sdk, /docs, /pricing, /benchmark.
• Krisp: krisp.ai/developers/voice-isolation (models, 15 ms, 16 and 32 kHz, SDK languages, integrations); krisp.ai/developers/pricing (Apply Now, no price); krisp.ai/blog/voice-isolation-2-5 (CPU, narrowband support).
• Pipecat Cloud: daily.co/pricing/pipecat-cloud (Krisp VIVA, 10,000 free minutes, $0.0015/min).
• Telnyx: telnyx.com/pricing/call-control (Noise Suppression - Krisp, $0.005 per leg per minute).
• ai-coustics: ai-coustics.com/pricing (tiers, under 30 ms, on-prem SDK); docs.ai-coustics.com/reference/sdk/models (Voice Focus 2.2, 16 kHz, SDK languages); docs.ai-coustics.com/guides/models (8 to 192 kHz, internal resampling).
• AssemblyAI: assemblyai.com/docs/streaming/voice-focus (modes, threshold, models); assemblyai.com/pricing (+$0.10/hr, $0.45/hr base, $50 credit); streaming quickstart (sample_rate matches your source).
• LiveKit: livekit.com/pricing (voice isolation models, included minutes, $0.0012/min); docs.livekit.io noise cancellation (where it runs, plugins, primary-speaker wording).
• NVIDIA: Speaker Focus effect (early access, 16 or 48 kHz, GPU throughput); AFX SDK user guide (platforms).
• Picovoice: picovoice.ai/docs/koala (on-device, platforms, AccessKey); Koala Python API (sample rate read at runtime); Pipecat Koala filter.
• Open source: DeepFilterNet (48 kHz, MIT or Apache-2.0); GTCRN (MIT, 48.2K parameters).

FAQ

What is the difference between speaker isolation and noise suppression?

Speaker isolation keeps one primary speaker and removes everything else, including other voices. Noise suppression removes non-speech noise and keeps all speech. In the table, Krisp VIVA, ai-coustics Voice Focus, AssemblyAI Voice Focus, LiveKit voice isolation, NVIDIA Speaker Focus and Anecho market isolation; Koala, DeepFilterNet and GTCRN are noise suppression. See what is speaker isolation.

Which speaker isolation SDKs run on CPU in my own process?

Of the products we could source, Krisp's server SDK (15 ms, CPU), ai-coustics' on-prem SDK (under 30 ms) and Anecho's SDK (15 ms, CPU) run in your own infrastructure. NVIDIA Speaker Focus publishes GPU throughput. AssemblyAI and LiveKit's slot run in their clouds.

Which speaker isolation products publish a price?

As listed on 2026-10-20: Anecho ($0.0010/min after 10,000 free minutes a month), LiveKit Cloud ($0.0012/min after included minutes), ai-coustics (from $135 a month billed annually) and AssemblyAI (+$0.10/hr). Krisp direct and NVIDIA do not publish a price; Krisp VIVA is resold at $0.0015/min on Pipecat Cloud.

Is Anecho better than ai-coustics?

It depends on the recogniser and the condition. On our live-conversation benchmark the pooled result is parity with faster-whisper base.en (−0.13 pp [−4.56, +4.47]), where no engine beats the untouched audio, and a v8_2 lead with Gemini Live input transcription (−6.47 pp [−9.89, −3.26]). v8_2 leads with a close-mic talker and a loud background; Quail VF 2.2 S leads with a distant talker, on pause suppression and on false barge-ins; Quail L leads on plain noise.

Have you measured Krisp or AssemblyAI?

No. Neither is on our benchmark, so this post makes no claim about their quality. Their own published figures are company claims.

Does speaker isolation always improve transcription?

No. On our 19-condition matrix the untouched audio scored 14.32% WER and every engine was higher pooled. Test on your own calls against a do-nothing control before switching isolation on for everything.