ASR leaderboard
Word error rate on speech recorded across a room: echo, distance and background noise. Lower is better.
Analysis
Where each model breaks down as the room gets harder, how accuracy trades against speed, and how the ranking moves. Charts follow the controls below.
Accuracy vs speed
Models on the Pareto front have no other model that is both more accurate and faster. Picked models get a teal ring.
Robustness across conditions
WER for each model from close-mic speech to measured, simulated, noisy and moving-talker conditions. Darker cells are worse.
Ranking dynamics
How the order changes from the easiest condition to the hardest, and as new models arrive. Press Play to step through.
Listen
Hear what the models hear: the same kind of speech from a close microphone, then across a room with echo and rising background noise.
The simulated room
Methodology
How FFASR builds its test conditions, scores models, and handles your data.
Cite FFASR
Copy this BibTeX entry.
Submit a model
Paste a Hugging Face model id. FFASR loads it with an automatic backend: Whisper, Granite speech, Cohere Transcribe, Wav2Vec2 and HuBERT CTC heads, SpeechBrain and most other ASR models work without extra setup.
- SubmitPaste the model id and your email.
- ReviewA moderator approves each new request, and any custom code, before it runs.
- EvaluateIt runs on one NVIDIA L4 against the hidden test set.
- ResultsScores appear on the leaderboard. Track progress with your job id.
Type a model id as author/name to get a link to its Hub page.
Check a job
Paste the job id you got when you submitted.
Up next
Models waiting for evaluation. Questions about a submission or a gated repo? Email contact@treble.tech.
Loading the queue…
FFASR by Treble Technologies. Models come from the Hugging Face Hub and run on Hugging Face Jobs against a hidden test set.
Questions: contact@treble.tech · Moderator tools