Local ASR (Speech-to-Text)

Comparative review of three local speech-to-text approaches tried across separate projects: Meta's omnilingual_asr, Microsoft's VibeVoice-ASR, and onnx-asr.

Evaluated 2026-08-11, covering work from 2026-08-01 through today.

What we needed

• Offline/local transcription — no cloud STT dependency
• Works without a baby-sat persistent server
• Tolerable on CPU, since the one 3090 in the house is usually pinned near 24GB by a standing ComfyUI process
• Good enough accuracy for both short voice notes and longer batch jobs (talks, multi-hour recordings)
• Minimal install/setup footprint

Approaches

Meta omnilingual_asr (omniASR_LLM_3B_v2)

Used in ~/asr-test (2026-08-02/03) for a large Hindi (hin_Deva) batch transcription job — a hand-built pipeline: slice audio into clips, build a manifest, transcribe in batches of 8 with resume-from-checkpoint support, merge translations.

Verdict: not adopted going forward. It worked, but it's a bespoke script around a 3B model, not a tool — no CLI, no reusable driver, 7.6GB of venv + intermediate artifacts for one job. Right choice for a big one-off multilingual batch job; wrong choice as a general-purpose default.

Microsoft VibeVoice-ASR (raw fallback)

Used in ~/code/vibevoice-transcribe (2026-08-01) to transcribe a set of conference talks. The intended path — a persistent vibevoice-server — is broken: it crashes on model load because transformers 5.14.1 dropped the eager-attention fallback VibeVoice's tokenizer encoder needs, and the server CLI exposes no way to force it. Workaround was run_raw_transcribe.py, which loads the model directly from the VibeVoice repo venv on every invocation (tries sdpa then eager), rather than talking to a long-running server.

Verdict: not adopted going forward. Functional once you build the workaround, but fragile — a broken server, a hand-rolled reload-per-run driver, 12GB footprint, and a transformers-version landmine that will resurface on any upgrade.

onnx-asr (CLI)

Reached for today to transcribe a WhatsApp voice note. Already installed as a uv tool — no project setup at all.

• whisper-base: garbled/incomplete output. Skip it.
• nemo-parakeet-tdt-0.6b-v2 with --vad silero: clean, correctly-segmented transcript on the first real attempt.

Runs CPU-only here — its isolated uv tool venv doesn't bundle cuDNN/TensorRT (unlike the two projects above, which pull nvidia-cudnn-cu12 via pip into their own venvs), so onnxruntime logs EP-load errors and falls back automatically. Not a functional problem, just slower: a ~90-second clip took a couple of minutes. Total footprint: ~526MB tool + ~2.8GB model cache, an order of magnitude lighter than either alternative above.

Verdict: recommended. One CLI invocation, no server to babysit, no crash history, good accuracy with the right model. Default going forward for anything short-to-medium; if a large batch job needs GPU throughput, install nvidia-cudnn-cu12 into a dedicated venv rather than relying on the uv tool isolation, and reconsider omnilingual_asr only if the content is heavily non-English (its multilingual coverage is broader than Parakeet's).

Housekeeping

~/asr-test (7.6GB) and ~/code/vibevoice-transcribe (12GB) were deleted 2026-08-11 — both fully wrapped, Gavin confirmed disposable.