OpenAI: Whisper 1

Whisper 1 is OpenAI’s general-purpose speech-recognition API model, documented as powered by the open-source Whisper V2 model. It accepts an audio file and returns text. The maker describes multilingual recognition, language identification and a separate speech-translation task. Transcription preserves the spoken language; the producer’s separate translation interface returns English text.

A useful application is preparing an interview, voice note or recorded conversation for later search and editing. Supply the completed recording and, when known, its input language. OpenAI documents 98 supported languages and notes that recognition accuracy varies by language. The singular language hint uses an ISO 639-1 code; the newer multi-language and keyword fields of other transcription models are separate controls.

The producer’s file-transcription flow accepts files up to 25 MB in MP3, MP4, MPEG, MPGA, M4A, WAV or WebM format. For a larger recording, use a compressed format or divide it into chunks within the limit. Preserve sentence boundaries and the order of the chunks so that the assembled transcript remains connected to the original recording. A video container supplies its audio for this task rather than a visual-description input.

A prompt can supply names, acronyms or recording-specific vocabulary and context from an earlier segment. Whisper’s prompt limit is 224 tokens, and the prompt should match the audio’s language. The maker states that Whisper does not follow instructions like a general-purpose text model. Use the prompt as relevant transcription context, then check important spellings against the recording rather than treating a terminology list as words the speaker must have said.

The producer documents JSON, plain text, SRT, VTT and verbose JSON transcription formats. Word and segment timestamps require verbose JSON; either or both granularities can be requested. The API reference notes that segment timestamps add no extra latency, while word timestamps require additional processing. These outputs support captioning and locating passages during editing; speaker-diarization annotations belong to a different specialized model.

Whisper 1 does not support file-transcription streaming. Its reference also documents a sampling temperature between zero and one; at zero, the model can increase temperature automatically when its decoding thresholds require it. This describes the producer’s processing controls rather than a promise that every recording produces an identical or perfectly accurate transcript.

Define the final artifact: a transcript for search, a subtitle file or text with word/segment timing. Keep the original recording, chunk order and relevant source time references with the result. A later text-editing step can correct a misspelling or prepare a summary, but compare changes with the audio so the speaker’s meaning is preserved. The recognized text becomes evidence for the next task rather than a generated answer to the spoken question.