AirCaps A5S Async v1 Benchmark - AISHELL-4 - cpCER: leaderboard
AirCaps A5S Async v1 Benchmark - AISHELL-4 - cpCER — AirCaps A5S Async v1 author comparison, November2026. Same 16kHz mono16bit WAV full recordings and native API defaults, with only exact model, language and diarization set. English language is en; Mandarin datasets use their native language. Meta switches from file to realtime above10min with same model/mode. Other documented request limits use fewest equal chunks, two-minute overlap and reference-blind stitching. Failed recordings are empty transcripts/all deletions. References and hypotheses receive OpenASR normalization plus ellipsis removal and split-numeral joining, per speaker runs, then MeetEval minimum-permutation error. Nvidia is BF16 self-hosted1xH100, temperature0,top_p1,seed0,reasoningoff and its published benchmark prompt; NVIDIA generation-loop guardrail retained. Model runs and native request settings are attached. AirCaps accuracy runNovember6; peersNovember1–2. This campaign is distinct from VoiceArena or another evaluator using the same corpus. Mandarin character-error metric, not word error; percentage values retain native cpCER units. Metric: cpCER (%). Source: https://github.com/aircaps/a5s-async-v1-benchmark
Metric: cpCER (%). Source: github.com. Saturation forecast: Rough model projection: around 2026. 9 models tracked. Updated 16 Nov 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | AirCaps A5S Async v1 | 18.43 | |
| 2 | AssemblyAI Universal-3.5 Pro | 19.27 | |
| 3 | Muse Voice Transcribe 1.0 | 27.9 | |
| 4 | Nemotron 3.5 Transcribe | 31.11 | |
| 5 | Scribe v2 | 32.34 | |
| 6 | MAI-Transcribe-2 | 35.2 | |
| 7 | Gemini 3.5 Transcribe | 54.77 | |
| 8 | Grok Voice Transcribe 2.0 | 58.23 | |
| 9 | Deepgram Nova 3 | 68.19 |
Interactive version: theaggregate.ai/benchmark?slug=aircaps-a5s-async-v1-benchmark-aishell-4-cpcer · How It Works · Data refreshed daily, snapshot 2026-11-16.