Feature Proposal: Add FunASR Audio Transcription for Searchable Audio Archives

[LauraGPT]

ArchiveBox archives web content including audio/video files. Adding speech-to-text for archived audio content would make it searchable alongside text content. FunASR (17.8K+ stars, https://github.com/modelscope/FunASR) provides:

• SenseVoice: Ultra-fast multilingual ASR (50x faster than Whisper-large)
• Paraformer: Production-grade ASR with timestamps and punctuation
• OpenAI-compatible API: POST /v1/audio/transcriptions

Use case: When archiving pages with embedded audio (podcasts, interviews, meeting recordings), FunASR can transcribe the audio content and add it to ArchiveBoxes searchable index. This makes audio content as discoverable as text content.

Since ArchiveBox is self-hosted and FunASR also runs locally, they integrate naturally without external API dependencies.

Would adding FunASR transcription for archived audio be useful?

[pirate]

Yes I'm down, feel free to add an optional plugin for it that's disabled by default. (any plugins that depend on external servers or APIs should be disabled by default)

[LauraGPT]

Sounds good. I’ll scope this as an optional plugin that is disabled by default and does not add any dependency to ArchiveBox core.

The useful first version would be: configurable local command or HTTP endpoint, write transcript sidecars/metadata, then let ArchiveBox index the transcript text. That keeps external servers and heavier ASR runtimes outside the default path.

[videodb-kal]

This seems aligned with the core need here: not just ASR, but making archived audio/video searchable by what was said. If you're evaluating FunASR or an OpenAI-compatible transcription endpoint mainly to index podcasts and other embedded media, you can replace the transcribe → chunk → embed → vector-search path with one hosted indexing step plus a search call that returns matching timestamps.

# Transcribe a video and semantically search what was said — VideoDB
# pip install videodb   (docs: https://docs.videodb.io)
import videodb
from videodb import SearchType, IndexType

conn = videodb.connect()  # reads VIDEO_DB_API_KEY from env

video = conn.upload(url="https://www.youtube.com/watch?v=WDv4AWk0J3U")

# One call replaces whisper + chunking + embeddings + a vector store:
video.index_spoken_words()

results = video.search(
    query="find spoken content inside archived audio/video so it becomes searchable alongside text content",
    search_type=SearchType.semantic,
    index_type=IndexType.spoken_word,
)
shots = results.get_shots()
for shot in shots:
    print(f"{shot.start:.1f}s - {shot.end:.1f}s  {shot.text}")

# Instant playable clip of just the matching moments:
if shots:
    stream_url = video.generate_stream(timeline=[(int(s.start), int(s.end)) for s in shots])
    print(stream_url)

That gives you a way to surface spoken matches alongside the rest of ArchiveBox's text search without having to wire up the transcript segmentation and retrieval layer yourself. It is a hosted service though, so it would only make sense if a third-party media indexing API is acceptable for this plugin.
(I help build VideoDB — sharing because it fits this.)

Docs: https://docs.videodb.io/?utm_source=forge1&utm_medium=github&utm_campaign=b2e2c841-d719-4590-899d-6cbb961d6b9e&utm_content=A&utm_term=transcribe-search