Fish Audio Speech to Text Multispeaker Transcription

Fish Audio speech to text turns a recording into a labelled transcript: who spoke, what they said, and how each line was delivered. Transcription runs on the same credits as generation — one credit covers about 1,000 characters — and it is included in every pack, so a $9 Starter pack is enough to transcribe interviews, calls or raw takes before you decide where the project goes.

What is included in a Fish Audio transcript?

Four things arrive with each transcription:

• Text of what was said, in the language it was spoken.
• Speaker separation, so a recording with two or more voices comes back as labelled dialogue instead of one unbroken paragraph.
• Emotion tags on lines where the delivery matters — laughter, a raised voice, a pause — in the same bracket format the text to speech engine reads.
• Timing context you can use to cut a script to the beats that actually matter.

That combination is the reason transcription sits next to generation in the same product rather than in a separate tool: the transcript you produce is a script you can immediately re-record.

How does multispeaker transcription work?

Interviews, panels, podcasts and support calls share one problem: content is only useful if you can attribute it. A wall of text without speaker labels is unsearchable and unquotable.

Multispeaker transcription separates the voices present in the audio and labels the turns. The practical benefits are immediate:

• Editing a dialogue script becomes line-by-line work rather than guesswork.
• Quote attribution for marketing or journalism is verifiable.
• Search and QA tooling can filter by speaker — "what did the customer say" stops returning the support agent.

The output is also the input for regenerating audio. Once turns are labelled, you can re-voice a single speaker's lines without disturbing the rest of the recording, which is the workflow described in the voice cloning guide.

What do emotion tags add?

Flat transcripts lose the part of speech that carries meaning. Bracket tags preserve it. A support call transcript where a line is tagged as hesitant reads differently from one where the same words are tagged as confident, and that difference survives into whatever you build on top of the transcript — a report, a highlight reel, or a re-recorded narration.

It also removes a manual step from generation. Because the tags are already in bracket form, a transcript can become a script for text to speech with the pacing and emotion direction intact.

Which languages are supported?

Speech to text covers the same 30+ languages as the voice engine, including English, Japanese, Korean, Chinese, French, German, Arabic and Spanish. Mixed-language recordings are common in international teams, and keeping transcription and generation in one product means a French segment inside an English call does not force a tool switch.

Language coverage also compounds in the other direction: a transcript in one language can be regenerated in another with the same voice, which is how series stay recognisable across markets.

How does transcription fit with the rest of the workflow?

Three paths matter in practice:

1. Transcript → publish. Clean up the text, keep the tags, and re-record with a consistent voice using the text to speech guide.
2. Transcript → analysis. Feed speaker-labelled, emotion-tagged text into search, compliance review or customer research.
3. Transcript → long-form. Use it as the outline for narration-length content; the Studio guide covers chapter-level control for longer pieces.

In all three the transcript is the artefact you own. The audio can be regenerated; the text is the source of truth.

What does transcription cost?

Transcription spends credits the same way generation does, one credit per roughly 1,000 characters of transcript:

Because credits are one-time purchases that stay valid for 12 months, transcription is not a subscription decision. A quarter of interviews can be processed in one month and the remaining credits saved for the next production cycle.

What are the practical limits?

Length is the constraint worth planning for. Two habits keep transcription reliable:

• Split long recordings at natural boundaries — a call, a session, a chapter — rather than transcribing a six-hour file in one pass.
• Keep the raw audio alongside the transcript. If a passage is unclear later, the original is still there to re-process rather than to reconstruct from memory.

For generated audio, the same discipline applies in reverse: generate in paragraphs, keep the script, and re-render only what changed.

How do you prepare audio for a clean transcript?

Transcription is only as good as the recording it receives. A few minutes of preparation saves an hour of correction:

• Reduce background noise at the source. A microphone close to the speaker beats any amount of cleaning applied afterwards.
• Keep one speaker per channel when you can. Even when the model separates voices, clean separation improves labelling.
• Avoid overlapping music beds. Speech under a music track loses the quieter syllables that carry meaning.
• Note the context before you upload. Order numbers, product names and jargon are exactly the terms that need a spell-check pass afterwards, so keep a reference list beside the transcript while you review it.
• Cut long silences. They cost characters without adding content, and a transcript reads better without minutes of room tone between answers.

Afterwards, do one editing pass before the transcript becomes a script. Fix names and numbers first, keep the emotion tags intact, and leave the spoken syntax alone: the way people actually talk is why a regenerated read sounds human rather than read aloud.

Where should you start?

The fastest path is to transcribe one real recording you already care about, then decide what to do with it:

• Buy a $9 Starter pack and transcribe the sample.
• Read the text to speech guide if the plan is to re-voice the transcript.
• Read the voice cloning guide if a specific speaker's voice has to stay consistent.

Transcripts are the cheapest form of first-hand material you can build on. Start with one, and the rest of the workflow becomes obvious.

Frequently asked questions

What is included in speech to text?

A text transcript of the audio, with speaker separation for multi-voice recordings and emotion tags that mark how a line was delivered.

Do I need a separate transcription plan?

No. Transcription is part of the same credit system as text to speech — every pack includes it, including the $9 Starter pack with 1,000 credits.

Can it tell who said what in a meeting or interview?

Yes. Multispeaker transcription separates voices in one recording, so interviews, panels and support calls come back as labelled dialogue rather than one continuous block of text.

Which languages can be transcribed?

The same 30+ languages the platform speaks, so a recording does not have to change tools just because the participants switched language.

What should I do with the transcript afterwards?

Most teams edit it into a script and regenerate the audio with a consistent voice, or feed it into search and QA tooling — the text to speech and cloning guides cover both paths.