Fish Audio: Transcribe 1

Fish Audio: Transcribe 1 is a speech-to-text model from Fish Audio, listed on api14 under the transcription category. It is the provider's dedicated transcription offering, distinct from any text or multimodal generation work, and it exists purely to turn recorded speech into written text. If your application needs to convert audio into a usable transcript and you are already routing other requests through an OpenAI-compatible gateway, this is the model to evaluate for that job.

The model accepts audio as its only input modality and returns a transcription as its output. It performs automatic language detection, so you do not have to tell it in advance which language is being spoken before it can process a clip. When alignment details are requested, it can return timestamped word-level segments rather than one undifferentiated block of text. That word-level timing is the feature that matters most for anything downstream of plain reading: it lets you map each word back to a point in the audio. The model metadata does not publish a context length in tokens, so treat the audio itself, not a token budget, as the thing you are sizing your requests around.

Because the output can carry word-level timestamps, several practical use cases follow directly. Subtitle and caption generation benefit from segment timing, since a caption track needs to know when each phrase appears. Transcript editors and note-taking tools can highlight the audio as it plays, or let a user click a word to jump to that moment. Search and indexing over recorded meetings, interviews, lectures or support calls becomes possible once the speech is text, and word alignment improves the precision of those hits. Language detection also helps when you receive mixed or unknown-language audio and want a transcript without a separate classification step first.

You call the model through api14's OpenAI-compatible API using the model id "fish-audio/transcribe-1". Because the interface follows the OpenAI convention, existing clients and SDKs that already point at an OpenAI-compatible base URL can send requests without a bespoke integration. The model lists no supported request parameters, so plan on the standard request shape for an audio transcription call and check the request schema for the fields that apply to audio input. When you need alignment output, request the alignment details so the response includes timestamped word-level segments instead of plain text. The metadata also records a max output value of zero, which means you should not rely on the published metadata for an output ceiling and should validate response sizes against your own workload.

On fit: this model is scoped to transcription only. It does not generate text, answer questions about the audio, translate, summarize or produce any output beyond the transcript, so a task like "summarize this meeting" needs a separate text model fed with the transcript. The absence of published supported parameters and a token context length means you cannot tune sampling or reasoning behavior here, and very long recordings should be tested against your own pipeline rather than assumed to work in one pass. If your requirement is speech in, accurate timestamped text out, with automatic language detection, this model is aimed squarely at that.