Summary

Add word-level timestamp extraction to the crispasr backend. The CrispASR C library (libgocraspasr) already exports word accessor functions, but the Go wrapper did not bind or call them.

What Changed

C shim (backend/go/crispasr/cpp/crispasr_shim.{h,cpp})

Added wrapper functions for two families of CrispASR word accessors:

Go wrapper (backend/go/crispasr/{main.go,gocrispasr.go})

• Registered the new C functions via purego.RegisterLibFunc.
• In AudioTranscription, after transcription, populates TranscriptSegment.Words:
  • Tries session-based per-segment word functions first.
  • Falls back to parakeet-specific global word list when session word count is 0 (and only on segment 0 to avoid duplication).

Why two function families?

CrispASR uses different backends internally. Whisper-style backends expose word timestamps per-segment through the session result (crispasr_session_result_n_words(result, seg_i)). Parakeet TDT models expose word timestamps globally through crispasr_parakeet_result_n_words(result) — the session-level word functions return 0 for these models.

Depends on

#10402 — the gRPC server wrapper must forward the Words field for the timestamps to reach the HTTP response.

Testing

Verified end-to-end on LocalAI v4.4.3 (Docker, RTX 3090) with parakeet-tdt-0.6b-v3 via the crispasr backend:

{
  "segments": [{
    "id": 0,
    "start": 0.16,
    "end": 10.48,
    "text": "Been playing lots of Nordic Souls recently. Enjoy the list, but looking for something graphically better? Things I want to prioritize.",
    "words": [
      {"start": 0.16, "end": 0.4, "text": "Been"},
      {"start": 0.4, "end": 0.72, "text": "playing"},
      {"start": 0.72, "end": 0.88, "text": "lots"},
      {"start": 0.88, "end": 1.04, "text": "of"},
      {"start": 1.04, "end": 1.6, "text": "Nordic"},
      {"start": 1.6, "end": 2.0, "text": "Souls"},
      ...
    ]
  }]
}

21 words returned with accurate timestamps matching the TDT duration head output.