AirCaps Voice Lab
Talk to sales Start free

How we teach AI to listen.

Technical reports from the AirCaps Voice Lab: how each model is built, what we test it on, and every number behind the results.

LatestSpeech-to-text

A5S Async v1: the most accurate speaker-attributed transcription for real meetings.

How A5S Async v1 makes 29% fewer errors than the next-best model on five public meeting benchmarks, how it holds up as overlap and speaker count rise, and what it costs and how fast it runs, measured against frontier models from AssemblyAI, ElevenLabs, Microsoft, NVIDIA, Deepgram, Google and Meta.

Read the note
5 of 5public meeting benchmarks with the lowest error
21.74%cpWER averaged over those 5 benchmarks
29%fewer errors across 5 public meeting benchmarks

More notes.

Diarization-v1
12.03%average DER over 7 public datasets

Speaker diarization

Diarization-v1: who spoke when, across nine public datasets

Meetings, phone calls, video and broadcast conversations, evaluated against open-source and commercial systems. 42–70% lower error than the best commercial speech recognition model on each of seven benchmarks.

Read the note
Speaker ID-v1
3.11%EER under noise and music, from 17M parameters

Speaker identification

Speaker ID-v1: voice recognition that holds up in noise

A compact speaker identification model, evaluated on VoxCeleb1-O with environmental noise and music louder than the voice: 3.11% equal error rate, against 5.06% for TitaNet Large.

Read the note
A5S Small V2
26.27%WER, the lowest of 9 streaming systems

Streaming speech-to-text

A5S Small V2: streaming speech recognition for noisy, far-field audio

Real-time English speech-to-text evaluated on Mega-ASR, AMI, DiPCo and NOTSOFAR against 8 cloud streaming systems, with side-by-side recordings from the field.

Read the note

Open benchmarks.

Every result can be rerun: dataset recipes, provider adapters, raw API responses and scores are public.

A5S Async v1 benchmarkFive public meeting benchmarks: recipes, adapters, raw outputs and scores.GitHub AirCaps datasetsEvery evaluation set we publish, in one place.Hugging Face Diarization benchmarkNine public datasets, scored the same way for every system.GitHub Streaming ASR benchmarkFour real-world corpora and 9 streaming systems.GitHub Mega-ASR noise setNoisy, far-field speech for streaming evaluation.Hugging Face VOX1O-MUSANVoxCeleb1-O with environmental noise and music, for speaker ID.Hugging Face

Try the models behind the research.

Run them on your own audio. Every account includes 5 free hours.

Start free See the models