Feature: Add FunASR as audio transcription component (13x faster, speaker diarization)

[LauraGPT]

Summary

Haystack currently supports Whisper (local + remote) for audio transcription. FunASR (16.5K stars, MIT) would be a valuable addition as an alternative audio component — significantly faster, with built-in speaker diarization and emotion detection.

Why FunASR for Haystack pipelines

	WhisperLocal	FunASR
GPU speed	13x realtime	170x realtime
CPU speed	❌ Too slow	17x realtime
Speaker diarization	❌	✅ built-in
Emotion detection	❌	✅
Timestamps	✅	✅ word-level
Languages	57	50+
API mode	OpenAI-compatible	✅ OpenAI-compatible

For RAG pipelines that index audio/video content, the speed difference is critical — processing a corpus of meeting recordings or podcast episodes is 13x faster.

Speaker diarization is also valuable for RAG: knowing who said something enables filtering by speaker in retrieval.

Integration approach

FunASR already exposes an OpenAI-compatible API:

funasr-server --device cuda
# POST /v1/audio/transcriptions

So a FunASRTranscriber component could follow the same pattern as RemoteWhisperTranscriber, pointing to the local FunASR endpoint. Or directly via Python:

from funasr import AutoModel

model = AutoModel(
    model="iic/SenseVoiceSmall",
    vad_model="fsmn-vad",
    spk_model="cam++",
    device="cuda"
)
result = model.generate(input="meeting.wav")
# Returns: text with speaker labels, word timestamps, emotion tags

Use cases in Haystack pipelines

1. Audio RAG: Index meeting recordings → retrieve by content + speaker
2. Podcast search: Transcribe + index episodes for semantic search
3. Video understanding: Extract text from video files for multimodal RAG
4. Call center analytics: Transcribe + detect emotion for quality monitoring

Happy to contribute a FunASRTranscriber component if there's interest.

[Aakilahamed006]

This looks like a valuable addition to Haystack's audio processing capabilities. The built-in speaker diarization and emotion detection provided by FunASR could be particularly useful for audio RAG use cases, where speaker-aware retrieval can improve the quality of search and question answering over meeting recordings and podcasts.

I also like that FunASR provides both a Python interface and an OpenAI-compatible API, which seems to fit well with Haystack's existing component architecture.

I'd be interested in exploring an implementation of a FunASRTranscriber component, following a similar pattern to the existing Whisper transcribers. If the maintainers think this would be a good fit for Haystack, I'd be happy to work on a design and potentially open a PR.