Fish Audio

Fish Audio

This connector uses the Fish Audio developer API (https://api.fish.audio): multilingual text to speech with a large library of voices and custom voice clones, speech to text with speaker turns, voice model management and the API credit balance.

Setup (2 minutes)

1. Sign up at https://fish.audio and open Developers, Get started (https://fish.audio/app/developers/get-started/).
2. Click Get your API key; the key is shown only once. Paste it into FISH_AUDIO_API_KEY. It is sent as Authorization: Bearer.
3. New accounts have no API credit: text to speech with the model s2.1-pro-free works at once, every other model and transcription return 402 until credit is added under Developers, Billing (https://fish.audio/app/developers/billing/). The developer API credit is separate from the web app subscription.

Costs: fish_audio_text_to_speech costs $15 per million UTF-8 bytes of text with s2.1-pro or s2-pro (about $0.015 per 1,000 Latin characters; Chinese and Japanese characters take 3 bytes each), $50 per million with drama-3, nothing with s2.1-pro-free. fish_audio_transcribe costs $0.36 per hour of audio. Tell the user the cost of long texts or recordings and confirm first. Failed requests are not billed. All other tools are free.

Workflow

• Balance: fish_audio_get_api_credit (developer API credit in USD, the probe) and fish_audio_get_package (web app plan).
• Find a voice: fish_audio_list_voices with title, tag (male, female, young, old, narration, calm, deep ...), language (en, de, zh ...) and sort_by task_count for the most used voices; licensed true for voices cleared by their owners. Play sample_audio to judge a voice. self true lists the account's own voices. fish_audio_get_voice shows one voice; state trained means ready.
• Speak: fish_audio_text_to_speech with text and the voice id as reference_id. The audio file is attached to the result. Use model s2.1-pro (default) or s2.1-pro-free on a free account; format mp3 is smallest; speed 0.5 to 2 and volume in dB adjust delivery; language helps read numbers and dates. Write cues in square brackets in the text for emotion and delivery: [happy], [sad], [excited], [whispering], [laughing], [sighing], [break]. Very long texts produce large files: split them into chapters (about 10 MB of MP3 is the limit of one result).
• Transcribe: fish_audio_transcribe with audio_url, a public link to the file (up to 10 MB here; AnythingMCP downloads and uploads it). transcribe-1-pro handles up to 60 minutes and several speakers: the text carries markers such as <|speaker:0|>; set ignore_timestamps false to get word segments (seconds) and speaker_turns (speaker, text, start, end). Speaker labels only identify speakers inside one transcript.
• Clone a voice: fish_audio_create_voice with a title and voice_url, a public link to 10 to 30 seconds of clean speech from one speaker, plus its transcript in text when known. The new voice _id works at once as reference_id. Only clone voices the user has the right to use. fish_audio_update_voice renames a voice or changes its description or visibility (tags are edited in the web app); fish_audio_delete_voice removes a voice permanently.

Reading answers: ids are 32 hexadecimal characters. Dates are ISO 8601 UTC. credit and cumulative_top_up are US dollars as decimal strings. Voice lists return total (capped at 1,000 for library searches) and has_more; page with page_number.

Limits and errors: concurrent requests are limited by spending tier (5 below $100 prepaid, 15 from $100, 50 from $1,000); a long transcription holds a slot until it finishes. Errors return JSON {status, message}: 401 "Invalid token" (wrong key), 402 no API credit (or a paid model on a free account), 404 unknown voice id, 422 or 400 for invalid fields (e.g. a speaker hint with diarize false), 429 too many concurrent requests (retry shortly).

Out of scope: voice design (returns candidate audio as base64 text), real-time WebSocket speech and transcription, the voice agent platform, and image and video tools. Fish Audio also runs its own MCP server at https://api.fish.audio/mcp, which signs in with OAuth and bills the web app plan instead of the developer API credit.

Switched off at install: fish_audio_delete_voice; they cannot be undone, so an admin switches them on when the team needs them.