Baseline Verified Pair
Recognition · Current safest baseline · idle · No result
Details
Runs gpt-transcribe and gpt-4o-transcribe independently and only accepts speech where the recognizers agree.
Recording
Upload a new recording or open one you've already saved.
Load audio or videoUpload a new recording or choose one already in your saved files. Playback, recovery, audio effects and transcription stay together on this page.
Transcription setup
Pick the coverage and audio version, then run one model or all of them.
What to transcribe
The whole recording is processed in consecutive 25-second sections and combined in time order.
Audio version
Local model server
Paste the current trycloudflare.com address from your running model-server tunnel. Model-server options are disabled until a URL is entered; built-in transcription remains available.
Upload a new recording or choose a saved one above to begin.
Whisper-1 Compatibility
Recognition · Single legacy Whisper pass · idle · No result
Details
Runs whisper-1 at temperature 0. Useful as a different acoustic opinion, but a single result is never treated as confirmed.
Whisper-Sensitive Ensemble
Recognition · Three-model cross-check · idle · No result
Details
Compares the verified pair with whisper-1 and separates words into confirmed, possible, and unclear instead of merging disagreements.
Whisper Detector + ASR
Acoustic heuristic · Routes likely whisper windows · idle · No result
Details
Uses a clearly labeled acoustic heuristic to find faint/breathy windows, then sends only those windows through the verified pair.
WavLM Baseline ASR
Recognition · General speech checkpoint · idle · No result
Details
Calls WHISPER_WAVLM_ENDPOINT when configured, otherwise the model server below with /wavlm. This is a GENERAL WavLM CTC speech checkpoint, not a whisper-trained one.
WESPER
Voice conversion → recognition · Whisper-to-voiced conversion · idle · No result
Details
Calls WESPER_ENDPOINT when configured, otherwise the model server below with /wesper. The converted waveform is synthesized audio; recognizer agreement on it is not acoustic confirmation of the original recording.
PALF-Whisper-small
Recognition · F0-mask whisper fine-tune · idle · No result
Details
Runs the jankoko/PALF-Whisper-small checkpoint on your model server. Trained on wTIMIT whispered and normal speech. A single model's guesses are not confirmed speech.
PALF-F1-Whisper-small
Recognition · F1-mask whisper fine-tune · idle · No result
Details
Runs jankoko/PALF-F1-Whisper-small through the same model server. It shares its Whisper-small architecture and related training data with the other PALF variants.
PALF-LF-Whisper-small
Recognition · Low-frequency-mask whisper fine-tune · idle · No result
Details
Runs jankoko/PALF-LF-Whisper-small through your model server. Compare the result with the original audio before accepting any faint words.
SpecAugment-Whisper-small
Recognition · Whisper fine-tune with augmentation · idle · No result
Details
Runs jankoko/SpecAugment-Whisper-small through your model server. It shares Whisper-small's architecture with the PALF models, so agreement is not independent proof.