Fish Audio: Transcribe 1 Pro

Transcribe 1 Pro is Fish Audio’s speech-to-text model for short clips, multi-speaker conversations and long recordings. It takes recorded audio and returns a transcript, with optional timing and structured speaker turns. Its distinguishing features are speaker changes marked in the text and cues about emotion or vocal events, which preserve more of the recording’s conversational structure than a plain sequence of words.

For an interview, meeting or podcast, keep the complete conversation in one file so that returning speakers can retain consistent labels within the response. A useful review artifact is a transcript linked to the original recording, with turns displayed separately and a timestamp available for checking a quotation. A summary or action list is a subsequent application step: keep it separate from the transcript so the spoken source remains available for comparison.

The producer accepts recordings up to 60 minutes and describes processing long files with compressed audio such as MP3, Opus or AAC. Supported formats include WAV, MP3, AAC in M4A or MP4, FLAC, Ogg with Opus or Vorbis, WebM/Matroska and MOV. For a video file, it uses the first audio track. AIFF, CAF, WMA, AMR, AC-3 and raw headerless PCM are unsupported. One audio file is supplied per request; the result arrives after the whole recording has been transcribed, and a long recording can take several minutes. Clips shorter than about 0.08 seconds are rejected.

The producer response contains the full text, the recording’s duration in seconds and a segments array. Pro marks speaker changes inline with labels such as <|speaker:0|> and <|speaker:1|>; a label applies to the following text until the next marker. These are identifiers within one response, not verified names or identities that can be matched across requests. If the transcript has no speaker marker, the guide treats it as a single-speaker result. Keep any text before the first marker with the first turn rather than discarding it.

Word-level timestamps are optional. In the producer request, ignore_timestamps defaults to true; setting it to false asks for segments with text, start and end times in seconds. A segment is generally a word, or one character or a few in Chinese and Japanese. Segment text omits punctuation, speaker markers and event cues, and can normalise a spoken value differently from the full transcript. A word segment can have the same start and end time. An empty segments array can mean timestamps were not requested, no speech was found or timing was temporarily unavailable. Use the full transcript for readable wording and segments for alignment rather than assuming both strings match character for character.

The diarize setting defaults to auto. When timestamps are requested and diarize is not false, Pro also returns speaker_turns. Each turn contains its speaker label, text and start/end times, preserving punctuation and retained cues without the inline speaker marker. If word timing is unavailable, turn times can be approximate and cover the whole recording; adjacent turns are not necessarily contiguous or non-overlapping. Disabling diarize removes the structured turn list but leaves speaker markers in the full text. A known speaker count, or minimum and maximum counts, can guide longer-recording diarisation on a best-effort basis. The exact count cannot be combined with the count bounds, and count hints cannot be used with diarize disabled.

Emotion and vocal-event cues appear as bracketed text, for example [laughter], alongside the recognised speech. They are inferences from the audio, not words that a user needs to add before transcription. Pro’s tag_audio_events setting defaults to true; disabling it removes these cues from the transcript and turn text without changing the timing. There is no separate emotion field or fixed emotion enumeration. Preserve the annotated transcript when the sound of a remark matters, and use a cue-free version when the downstream artifact needs only the spoken words.

Language is detected automatically. An optional lowercase ISO 639-1 hint does not force translation or change Pro’s transcript. If the language cannot be determined, the returned language code can reflect the hint without verifying it against the audio. The producer reports one language rather than an inventory of every language in a mixed-language recording. Keep the original wording and language metadata together when preparing a searchable archive, especially where different parts of a conversation switch languages.

For captions, group consecutive timed words into readable SRT or WebVTT cues rather than producing one caption per word. Separate cues around pauses and choose their length and display time to suit the player and audience. Speaker-turn boundaries can keep different speakers in separate cues. Check for an empty timing result and for cues whose end must be later than their start. Caption formatting, names for anonymous speaker labels and any later summary are application artifacts built from the transcription result.

For a practical recording workflow, decide before transcription whether you need a plain readable transcript, annotated conversation turns or timed captions. Keep the audio identifier with the result, retain the returned full text when punctuation and cues matter, and use the structured timing fields for navigation. Pro’s long-recording and multi-speaker capabilities are distinct from the general short-recording Transcribe 1 model; their limits and controls should not be treated as interchangeable.