Research noteNovember 2026Speech-to-text

A5S Async v1: the world's most accurate speech recognition model for speaker-attributed transcription.

Word errors with speaker attribution, 5 public meeting benchmarksMacro cpWER (%) · lower is better

#1on all 5 public meeting benchmarks

29% lowererrors than the next-best model across 5 public benchmarks

Introducing A5S Async v1, our most accurate speech recognition model yet. On audio with several speakers, its speaker-attributed transcripts beat frontier models from AssemblyAI, ElevenLabs, Microsoft, NVIDIA, Deepgram, Google and Meta.

Across 5 public meeting benchmarks it makes 29% fewer errors than the next-best model, with the lowest error on every one.

A5S Async v1 is built for real-world speech: people interrupting and talking over each other, in a meeting room or on a call, for hours at a time. Our benchmark harness is open source, from dataset preparation to scoring and raw API responses.

Start free Talk to salesBenchmark code ↗API docs ↗

Lowest error on every public multi-speaker benchmark

We ran A5S Async v1 on 5 public datasets of natural meetings: 3 in English and 2 in Mandarin, 226 recordings and 59 hours in all. It has the lowest error on every one. On NOTSOFAR-1, the largest, it scores 16.72% cpWER against 24.28% for the next-best model.

cpWER by dataset% · lower is better

A5S Async v1
Other models
→ above 100%

All results

System	NOTSOFAR-1	AMI	CHiME-6	AISHELL-4cpCER	AliMeetingcpCER	Macro
AirCaps A5S Async v1	16.72	16.07	42.60	18.39	14.90	21.74
NVIDIA Nemotron 3.5 Transcribe	24.28	19.06	51.83	31.11	25.72	30.40
AssemblyAI Universal-3.5 Pro	34.10	25.85	51.32	19.27	29.50	32.01
Meta Muse Voice Transcribe	35.17	29.17	51.45	27.90	29.20	34.58
ElevenLabs Scribe v2	34.65	32.93	64.87	32.34	35.77	40.11
Microsoft MAI-Transcribe-2	35.94	42.76	64.06	35.20	30.03	41.60
xAI Grok Voice Transcribe 2.0	44.60	46.97	81.46	58.23	58.18	57.89
Deepgram Nova-3	58.84	53.50	90.90	68.19	62.02	66.69
Google Gemini 3.5 Transcribe	39.16	62.83	108.84	54.77	69.09	66.94

cpWER (%), pooled per dataset (Mandarin per character). Macro is the unweighted mean. † Sent in overlapping parts (per-request limit).

Supplementary datasets

We also ran 3 more multi-speaker datasets. Nemotron 3.5 Transcribe lists ICSI and DiPCo in its training data, so we leave it out of those 2. Agentic is a private dataset licensed from the Agentic Data Company. A5S Async v1 has the lowest error on all 3: 14.69% on ICSI, 32.09% on DiPCo and 8.65% on Agentic.

cpWER, supplementary datasets% · lower is better

Internal evaluation dataset

We recorded 7 English sessions in varied settings: 2–5 speakers, varying overlap, 13–59 minutes each and 3.4 hours in all. A5S Async v1 has the lowest error on 6 of the 7 and overall: 21.86%, against 30.87% for Meta and 31.84% for AssemblyAI.

cpWER per internal recording% · lower is better

A5S Async v1
Other models
→ above 100%

Data available on request.

Where the difference comes from

Most errors from the other models are deletions: speech that disappears when people overlap. On the English sets A5S Async v1 deletes 7.9% of reference words. AssemblyAI deletes 25.8%, ElevenLabs 18.8% and Nemotron 9.5%.

Error composition% of reference words · English sets

Deletions
Substitutions
Insertions

Real examples

Listen to clips from the public benchmark and follow each transcript against the human reference. Red words are wrong or extra. Struck-through words were missed.

Loading examples…

Lowest error at every level of overlap

On the English sets A5S Async v1 has the lowest error at every level of overlap. The lead is largest when half or more of the speech overlaps: 37% fewer errors than the next-best model.

cpWER by share of overlapped speech% · lower is better

A5S Async v1
Other models

By number of speakers

A5S Async v1 leads at every speaker count. With 5–8 speakers it makes 29% fewer errors than the next-best model, Nemotron (19.76% against 27.85% cpWER).

cpWER by speakers in the recording% · lower is better

A5S Async v1
Other models

On the frontier of accuracy, speed and cost

No model we tested is more accurate than A5S Async v1. Every model that is faster or cheaper makes 1.9–2.7 times as many errors. We co-designed a custom inference engine with our model to achieve extreme speed and efficiency.

End-to-end job speed

We timed each public API from upload to complete transcript, on hour-long meetings. One at a time, A5S Async v1 returns a meeting in 9.8 seconds (median). Only Deepgram (6.3) is faster. With 14 in flight it transcribes 3,739 audio hours per hour, behind only Deepgram (6,578).

Time to transcript, one hour-long meeting at a timeMedian seconds · lower is better

Throughput on hour-long meetingsAudio hours per hour · 14 requests in flight · higher is better

Accuracy against speed

On hour-long meetings only Deepgram is faster. On the same meetings it makes 2.7 times as many errors (67.17% cpWER, against 24.54%). No model is both faster and more accurate, so A5S Async v1 sits on the frontier, at its accurate end.

cpWER against speed, hour-long recordings (14 concurrent)Top right is best

Accuracy against price

A5S Async v1 costs $0.18 per audio hour with speaker labels. The 2 cheaper models make 1.9 and 2.7 times as many errors. Microsoft's $0.10 is an introductory price that ends December 31, 2026. Meta costs the same and makes 1.6 times as many errors.

Macro cpWER against list priceTop right is best

Pay-as-you-go list prices with speaker labels, November 2, 2026. *Microsoft: introductory price until December 31, 2026.

Built for meetings, close to the best on easier speech

Easier audio separates models far less, and A5S Async v1 stays close to the best. On DAI English, 2-person conversations, it scores 5.18% cpWER, behind Microsoft (4.12%) and ElevenLabs (4.39%). On 3 single-speaker Open ASR Leaderboard sets it averages 6.39% WER, 0.49 points behind Microsoft (5.90%), and it is best on VoxPopuli.

Error on easier speech% · lower is better

All results

System	DAI EnglishcpWER, 2 speakers	VoxPopuli	LibriSpeech clean	Earnings-22	Meansingle-speaker
AirCaps A5S Async v1	5.18	6.54	1.64	10.99	6.39
Microsoft MAI-Transcribe-2	4.12	7.10	1.91	8.69	5.90
NVIDIA Nemotron 3.5 Transcribe	5.72	6.76	1.44	9.88	6.03
ElevenLabs Scribe v2	4.39	7.51	1.82	9.33	6.22
AssemblyAI Universal-3.5 Pro	12.44	6.74	1.62	10.64	6.33
xAI Grok Voice Transcribe 2.0	15.48	7.67	2.11	9.74	6.51
Google Gemini 3.5 Transcribe	8.70	6.73	2.30	11.10	6.71
Deepgram Nova-3	16.89	8.61	2.57	15.00	8.73
Meta Muse Voice Transcribe	11.98	10.65	5.15	12.25	9.35

DAI English: cpWER (%). Single-speaker sets: WER (%), mean of the 3.

How we measured

cpWER

Words are grouped by speaker. Reference and model speakers are matched to give the fewest errors, and errors are counted over all reference words. So a word given to the wrong person counts as an error. Computed with MeetEval.

Same rules for every model

Every model gets the same 16 kHz mono file of the whole recording. We set only the model, the language and speaker labels: no speaker counts, vocabularies or prompts. The exception is Nemotron 3.5 Transcribe, which is prompted by design and gets NVIDIA's benchmark prompt. A5S Async v1 runs through its public API. There is no post-processing, and failures score as empty transcripts.

Same normalizer

We use the Hugging Face Open ASR Leaderboard normalizer, unchanged, plus 2 documented rules. Ellipses are removed like dashes, and split numerals are joined ("20 21" → "2021").

Each provider's most accurate model

We use each provider's best model for recorded audio. For AssemblyAI that is Universal-3.5 Pro, not the newer 3.6. Version 3.6 is a streaming model, and streaming is less accurate than batch.

Long recordings, split in good faith

Some models can't take a long recording in one request. We send those recordings as parts that overlap by 2 minutes and join them the way a customer could, without the reference. We align the shared words, link speakers by the words they share and remove the duplicate. This only helps those models. Scored strictly, with no speaker linked across parts, macro cpWER rises to 43.74% for Nemotron (from 30.40%), 49.77% for Microsoft (from 41.60%) and 82.00% for Google (from 66.94%).

Open source

The code, dataset recipes with checksums, provider adapters, settings, raw API responses, transcripts and scores are all public. Datasets are built from the official downloads, not redistributed.

Notes

Meta takes 10 minutes per file, so longer recordings went through its realtime endpoint, in parts past about an hour. Macro 34.58%, against 102.96% with 10-minute parts. Not in the speed tests.

Datasets. Nemotron lists ICSI and DiPCo as training data, and neither has a training split. Agentic: 4 online meetings licensed from the Agentic Data Company, machine-made references, available on request. Internal: our own recordings, referenced by third-party machine transcripts of close-talk mics, corrected for names and terms (2 checked by a person), never averaged into public results. DAI English: 48 2-person conversations, 2.0 hours (DAI-ASR-I18N English public split, mono mix). Single-speaker: 1,842 VoxPopuli, 2,620 LibriSpeech test-clean and 2,741 Earnings-22 clips. Recordings: CHiME-6 2, ICSI 3, Agentic 4.

Reading the numbers. Macro is the unweighted mean of the 5 main sets. Mandarin is scored per character. Breakdowns pool the 3 English main sets. cpWER exceeds 100% when a model inserts many words. Overlap comes from reference segments, which include short pauses, so it reads slightly high for everyone. Without CHiME-6, A5S Async v1 scores 20.02% at 30–50% overlap against 27.15% for the next-best model. Stitching can fail when parts share few words. That hit Google on 1 AMI meeting and 1 DiPCo session (274%), behind its 109.7% on hour-long meetings and 94.98% on DiPCo.

Speed. 8 English meetings of 45–61 minutes: AMI EN2002c, ICSI Bmr013, Bmr018 and Bro021, DiPCo S01, S03, S04 and S05. Latency: 1 meeting at a time. Throughput: 14 requests in flight, each meeting twice per round, 3 rounds, as requests in flight × audio ÷ latency. Untimed warm-ups first, median rounds reported. Providers ran from 1 AWS us-east-2 client (never the bottleneck), with no 429s or rate-limit waits. Split files go as parallel overlapping parts, timed to the last. Nemotron is self-hosted and not compared. Engine figures are internal (110 hours of meetings, full GPU load, start-up excluded) and can't be reproduced outside AirCaps.

Prices are pay-as-you-go list prices. Microsoft's $0.10 is introductory until December 31, 2026, and Microsoft hasn't published the price after that. AssemblyAI is $0.21 plus $0.02 for diarization. Deepgram is $0.0043 per minute. Google prices by token ($0.31 is its own estimate). Nemotron has no hosted price.

Scope. Each model ran once. Providers update models without notice. This report focuses on English, and the 2 Mandarin sets test generalization. A full multilingual report is coming soon.

What's next

Multilingual results

Our multilingual A5S Async model is already in internal use. This report covers English and 2 Mandarin sets. A full multilingual report on the same open benchmark is coming soon. Internal evaluations show the same gap to other providers in other languages as in English.

A5S Streaming

A streaming version of the model, in English and multilingual, with live speaker labels. On multi-speaker meetings it is only 2–4 points of cpWER behind A5S Async v1. Its lead over other providers' streaming models is larger than our lead in batch.

Try A5S Async v1 on your own audio

Start free

Every new account gets 5 free hours, then $0.18 per audio hour. Upload a recording in the Playground or use the API.

Start free API docs

Talk to sales

Send us a sample of your audio. We'll run an eval and show you real results.

Talk to sales

Building a product on speech, for meetings, contact centers, media, healthcare, legal or education? Talk to sales about volume pricing and private deployment.

Sources