Repository navigation
docs(tts): document the language field - #258
Conversation
Add the optional TTS `language` field (BCP-47 tag) to the non-generated pages: a Language section and a supported normalization languages table on the Text to Speech page, the WebSocket start event schema in asyncapi.yml, the fine-grained control overview, the fish-audio-api agent skill, and the changelog. The OpenAPI-generated reference pages pick the field up from the scheduled schema sync. Refs FISH-1399
|
Preview deployment for your docs. Learn more about Mintlify Previews.
💡 Tip: Enable Automations to automatically generate PRs for you. |
|
Warning Review limit reachedYou've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository. Next included review available in 52 minutes. View limit details
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info
📝 Walkthrough
Merge Risk: ⚪ Minimal · up to
|
Pull the current https://api.fish.audio/openapi.json (npm run update:openapi), which now has TTSRequest.language from platform-api#2433, so the check-openapi job passes and the generated API reference matches the Language docs in this PR.
Documents the optional TTS request field
language(FISH-1399, part of FISH-1394). It is live in production, andTTSRequest.languageis already in https://api.fish.audio/openapi.json (platform-api#2433).languageis a BCP-47 tag of the text (en,es-MX,zh-CN,zh-TW). It is used only by text normalization and word-timestamp alignment, not by the speech model. Omitted, empty orautokeeps per-chunk detection. Tags without rules skip normalization, with no error. The official SDKs don't expose it yet, so examples use raw HTTP.features/text-to-speech.mdx): new Language section with a curl example and the behavior (default, Spain vs Mexican Spanish,zh-TW, unsupported tags, mixed-language text,normalize: false, 400 validation, which routes accept it). New Supported normalization languages table: 13 rows, 42 tags. Spain/Mexico Spanish and Simplified/Traditional Chinese are separate rows, and anesvses-MXreading example follows the table.api-reference/asyncapi.yml):languageproperty onWebSocketTTSRequest, which applies to the whole session. Also fixes thenormalizedescription, which still said English and Chinese only.languagewhere normalization is discussed.languagerow, a correctednormalizerow, a session note for the WebSocket start event, and a 400 entry.Not touched:
api-reference/endpoint/openapi-v1/*andapi-reference/openapi.json. Those update through the scheduled schema sync. The open sync PR #251 was last refreshed at 03:06 UTC today and doesn't havelanguageyet, so the next 03:00 run should add it. SDK reference pages and the compat/ElevenLabs pages are also unchanged, because the compat gateway still ignoreslanguage_code.The
check-openapijob compares the committedopenapi.jsonagainst the live schema. It fails on every open PR until the sync PR is merged, so it isn't caused by this change.Need help on this PR? Tag
@codesmith-botwith what you need.Summary by CodeRabbit
languagefield for HTTP, SSE, and WebSocket requests, including its effects on text normalization and word timestamps.