Whisper
nigiri.whisper
A Docker Compose file for Whisper alone, generated from the catalog and checked by the safety inspector. No account needed.
Speech recognition service wrapping OpenAI's Whisper models, transcribing and translating audio across many languages with timestamps, served over an HTTP endpoint.
Whisper is an encoder-decoder model trained on a large multilingual corpus, which is what lets it transcribe accented speech, background noise and code-switching more reliably than earlier open recognition systems. It detects the spoken language automatically, emits segment and word-level timestamps, and can translate non-English speech directly into English text in one pass. Model sizes range from tiny to large, trading accuracy against memory and speed, and faster reimplementations of the same weights are commonly substituted for throughput. This deployment exposes the model as an HTTP service accepting uploaded audio and returning transcripts in several formats.
You know it worked when
- The service reports ready and its API documentation loads.
- An uploaded audio file returns a transcript.
- The detected language matches the audio's actual language.
- Segment timestamps align with the audio when checked against playback.
- A non-English file returns English text when the translate task is requested.
Known sharp edges
- The chosen model size is downloaded on first start, so the first request after deployment waits on a transfer of up to several gigabytes.
- CPU transcription of the larger models runs slower than real time, meaning an hour of audio can take longer than an hour to process.
- The model hallucinates plausible text during silence and music, producing confident transcripts of speech that never occurred.
- Uploaded audio frequently contains confidential conversation, so the endpoint needs authentication in front of it and a deliberate retention policy.