Generate MP3 speech with OpenAI TTS through RunAPI. Use for one-off speech generation or application integration. Prefer the RunAPI CLI for one-off requests and the target-language SDK for production integration.
Design & media
fish-audio
Try itCreate account-owned Fish Audio voice resources, attempt to reuse their IDs, or generate MP3/WAV speech through RunAPI. Use for voice resource management, one-off speech generation, or application integration. Prefer the RunAPI CLI for one-off requests and the target-language SDK for production integration.
What it does
Create account-owned Fish Audio voice resources, attempt to reuse their IDs, or generate MP3/WAV speech through RunAPI. Use for voice resource management, one-off speech generation, or application integration. Prefer the RunAPI CLI for one-off requests and the target-language SDK for production integration.
The skill document
Fish Audio on RunAPI
Choose route
- For a one-off artifact or result, use the registered
fish-audioservice in therunapiCLI. If the installed command catalog does not list it, stop and report the missing service instead of inventing a command. - For an app, backend, worker, library, webhook pipeline, or production codebase, go directly to Integrate with SDK. Never shell out to the CLI as the production runtime.
Discover contract
Authenticate, then inspect the installed command catalog and the selected operation's current contract:
runapi auth status > auth.json
jq -e '.authenticated == true' auth.json
runapi fish-audio --help
runapi fish-audio --help
curl --fail --location https://runapi.ai/docs/api/fish-audio/.md --output contract.md
If authentication is false, stop before submitting. Ask the user to provide a valid RUNAPI_API_KEY, or import a user-provided key from stdin with runapi auth import-token --token -; use interactive browser login only when the user explicitly requests it. Choose `` only from service help. Treat command help as authoritative for the installed operation, model, and top-level field roster. Treat its API Reference as authoritative for the complete request schema, nested fields, conditional rules, task behavior, and response variants. If the two surfaces disagree, stop and report the contract mismatch instead of guessing.
Build request
Create request.json as valid JSON using only fields accepted by the discovered operation contract. For the chosen model and values, evaluate every applicable conditional rule as a set: satisfy every required field, omit every forbidden field, and stop on unresolved contradictions.
Traverse nested objects and arrays before execution. Close every relationship stated by the discovered contract, including uniqueness constraints and cross-references between nested values.
For a discovered local media input, including file-typed fields and top-level media URL fields, put an agent-readable local file path directly in request.json. The CLI consumes file fields as declared and uploads local paths in top-level media URL fields. Use runapi files create only when the user needs a reusable URL, provides Base64, or the discovered contract explicitly requires a separate upload.
Semantic invariants:
- For voice resources, select
create-voice,list-voices, orget-voicefrom service help. Create requiresnameandsource_audio_url; get requires the RunAPIvoice_id; list accepts optionalpage_numberandpage_size. - Use a voice only after its state is
trained; a returnedvoice_idis a best-effort reference and may stop working later. - The
referencesaccepted bytext-to-speechapply only to the current request; they do not create a voice resource. - This version has no update, delete, revoke, or voice-library management operations. Do not assume a voice ID remains available or make claims about how source audio is used for training.
Validate the file before sending it:
jq empty request.json
Execute
Submit exactly once and save the complete synchronous response:
runapi fish-audio --input-file request.json > result.out
Do not add --async or call runapi wait. The operation completes in this response; preserve its exact JSON, text, subtitle, or other discovered response format.
Verify
A success status is not the deliverable. Read and validate the complete response according to the discovered result contract. Preserve the complete non-media result in the exact requested format, including JSON, text, SRT, or VTT.
For every requested media deliverable listed anywhere in the response, download all of them rather than returning only the first URL. Before downloading, derive its expected MIME type or family from response metadata when present, then the selected output format, then an unambiguous result field such as videos, images, or audios in the API Reference. The Catalog-declared fallback families for this skill are audio/*. Stop only when no single expected type or family can be established from those sources.
For every downloaded file, require both a non-empty file and the expected MIME type or family:
curl --fail --location --output
for file in ; do
expected_mime=
test -s "$file"
[[ "$(file --brief --mime-type "$file")" == $expected_mime ]]
done
Do not report completion when any requested deliverable is missing, empty, or has an unexpected MIME type. Record Skill Conformance separately from Task Outcome so a service failure does not hide whether this recipe was followed.
Recover or stop
- Correct a request shape at most once, and only when the discovered contract or returned validation error identifies the correction.
- Retry a transient transport failure at most once, and only when evidence confirms that no task was created, no billing occurred, and retrying is safe.
- On a terminal RunAPI or service failure, preserve the task/error evidence and stop. Keep the selected model and capability, and do not submit another paid request without user authorization.
- If the contract is missing a fact required to build or verify the request, stop and report the contract gap. Do not turn a product defect into a permanent skill workaround.
Integrate with SDK
Use this route only for application or production-code integration. Open the current RunAPI SDK reference below, select the package for the target language and Fish Audio, and confirm its install command, client methods, request types, response types, and error classes before coding. Build the request from the same discovered product contract and apply the same deliverable verification and stop rules. Do not invoke runapi as a subprocess from production code.
References
- Model overview, pricing, and rate limits: https://runapi.ai/models/fish-audio.md
- Provider overview: https://runapi.ai/providers/fish-audio.md
- Full model catalog: https://runapi.ai/models.md
- SDK integration: https://github.com/runapi-ai/fish-audio-sdk
Variants
Related skills
Generate speech, dialogue, and sound with ElevenLabs through RunAPI. Use when the user asks an agent to create speech, dialogue, or sound effects, or transcribe audio with ElevenLabs. Default to the RunAPI CLI for one-off generation; use SDKs only when the user is integrating RunAPI into an app or backend.
Fish Speech AI 配音工具 — 声音克隆 + 情绪分析 + 多段合成。用于短剧/AI剧配音制作,支持参考音频克隆音色、智能情绪识别、批量台词生成、音色库管理。
Transcribe uploaded audio through RunAPI with an OpenAI-compatible API. Use for one-off transcription, subtitle output, multilingual hints, or application integration. Prefer the RunAPI CLI for manual requests and the target-language SDK for production integration.
Use when the user asks to generate an image, create a video, make music, create audio, synthesize speech, submit a media task, poll a task, or produce genera...
Generate multi-speaker speech with Gemini TTS through RunAPI. Use when the user asks an agent to synthesize dialogue or integrate Gemini TTS. Use the RunAPI CLI for one-off generation and the language SDK for application integration.