Fish Audio: Transcribe 1

Transcribe 1 is Fish Audio’s general speech-to-text model for short recordings. It accepts an audio file and returns the recognised words as text, with the recording’s duration, optional timed segments and language information when known. This model is useful for a short voice note, a spoken instruction, an excerpt from a recording or an audio clip that needs a readable text representation.

An illustrative voice-note workflow keeps the recording together with its full transcript and an identifier for the original clip. A later application step can search that text, organise the note or develop a summary while preserving the spoken source. For a command, the transcript supplies recognised words to the application; any action is a separate interpretation and execution step. For an interview excerpt, record the clip’s position in the complete audio so a quotation can be checked at the right location.

The producer accepts WAV, MP3, AAC including M4A/MP4, FLAC and Ogg containing Opus or Vorbis. Browser recordings in WebM should be converted to Ogg/Opus, MP3 or WAV for this model; Pro documents native support for additional containers. Supply one audio file per request. The base model’s documented request allowance is up to 50 MiB, with MP3 and Opus files kept under 25 MiB. File size and recording length are different considerations: a compressed file can fit the upload allowance while still taking too long to process.

Transcribe 1 is designed for short recordings. The producer recommends Pro for audio longer than a few minutes; a long base-model request can fail if it exceeds the processing-time limit. When preparing a longer source for short-clip transcription, preserve each clip’s starting offset and confirm that the returned transcript covers the complete clip before combining the results. Add the original offset to local segment times to keep navigation connected to the full recording. The base model’s guidance should not be interpreted as Pro’s separate 60-minute allowance.

The producer response includes text, duration and segments. Duration is expressed in seconds, and the readable full transcript should be kept as the primary wording of the clip. Timestamps are optional: ignore_timestamps defaults to true in the producer API, while false requests word-level segments with text, start and end times in seconds. Requesting alignment adds processing time. An empty segments array can reflect skipped timing, no speech or temporarily unavailable alignment, so it should not automatically be interpreted as an empty full transcript.

Segments are generally words, or a character or a few in Chinese and Japanese. Their text omits punctuation and can normalise a value differently from the full transcript; it is not a character-for-character copy of the readable response. A segment can have the same start and end time. Use segments for finding a position in the recording or preparing timed display, and use the full text when the wording and punctuation matter. Keep those two representations distinct in the resulting artifact.

Language is detected automatically, whether or not an optional language hint is supplied. The producer documents lowercase ISO 639-1 codes for hints and returns a language code for application logic together with a language name for display when known. The response reports one language rather than every language in a recording that switches between languages. Keep the source wording and available language metadata with the result instead of treating the transcription request as a translation operation.

For a caption workflow, group consecutive timed segments into readable cues rather than showing one word at a time. A useful output is an SRT or WebVTT file linked to the original clip, with cue boundaries that fit pauses and a manageable amount of text on screen. Check for missing timing and ensure each cue’s end follows its start. Caption grouping and file formatting are application steps built from the returned alignment, and their display choices should suit the player and audience.

Transcribe 1’s documented controls are audio, a language hint and the timestamp choice. Pro’s speaker markers, structured speaker turns, emotion-event controls and speaker-count hints are separate features; they are not part of this base model’s documented response contract. Choose the required artifact before processing: plain readable text for a note, or text together with alignment for playback navigation and captions. Preserving the clip identifier, transcript and any available timing makes that result useful for the next application stage.