Feature request: Add self-hosted STT input with FunASR/SenseVoice

Motivation

Rasa users building private or on-premise voice assistants need a way to send recorded utterances to a self-hosted speech-to-text service. A provider-neutral OpenAI-compatible transcription contract could support FunASR/SenseVoice deployments without adding a heavyweight model runtime to Rasa itself.

Current architecture

This is not currently a provider-only change. On Rasa main, TwilioVoiceInput receives Twilio's completed SpeechResult text. Its response uses Twilio <Gather input="speech">, so raw microphone/call audio never reaches Rasa. The repository also has no ASREngine interface or generic voice/ASR configuration to implement against.

Smallest viable scopes

A maintainer decision is needed between two materially different features:

1. Generic recorded-audio input channel

   • Accept a multipart audio upload at a new Rasa input-channel route.
   • Forward the file to a configurable POST /v1/audio/transcriptions endpoint.
   • Require at least { "text": "..." }, then create the normal Rasa UserMessage.
   • Keep endpoint URL, optional bearer token, model, language, and timeout in server-side channel credentials.
   • This is the smaller, provider-neutral slice, but it does not replace Twilio phone-call transcription.

2. Twilio Media Streams plus pluggable STT

   • Add a WebSocket media route and call/session lifecycle.
   • Decode or forward Twilio's 8 kHz mu-law audio, define utterance boundaries, and handle disconnect/retry behavior.
   • Send completed audio to the configured transcription provider and preserve the existing tracker/output-channel flow.
   • This is a substantially larger phone integration and should not be represented as a simple FunASRASREngine class.

Backend scope

• An OpenAI-compatible endpoint can point to a compatible self-hosted FunASR/SenseVoice deployment or adapter; no cloud API is required.
• Direct in-process FunASR support should be separate because it adds model downloads and large optional runtime dependencies.
• SenseVoiceSmall supports Chinese, Cantonese, English, Japanese, and Korean. Other FunASR checkpoints have different language, timestamp, and hardware characteristics.
• FunASR and SenseVoice repository source is MIT-licensed; model weights follow their individual model cards.

Acceptance criteria for the smaller HTTP slice

• Mocked tests for authenticated and unauthenticated requests, config parsing, model/language forwarding, timeout, non-2xx response, malformed JSON, and missing text.
• Temporary audio is bounded and cleaned up on both success and failure.
• Provider errors are sanitized before logging or returning them to clients.
• Documentation contains one generic endpoint example and one tested FunASR/SenseVoice-compatible deployment example.

References

• FunASR: https://github.com/modelscope/FunASR
• SenseVoice: https://github.com/FunAudioLLM/SenseVoice
• SenseVoiceSmall model card: https://huggingface.co/FunAudioLLM/SenseVoiceSmall