Models and runtimes

This is the implementation catalog for the current checkout. uv run stt models is authoritative for accepted model IDs, defaults, and download estimates because those values can change with the code.

List, download, and select

Listing models does not download weights. The Offline column verifies whether each selectable model is complete in the active cache:

uv run stt models
uv run stt models --backend hf

Download one model through its backend's normal cache, then select that same ID for transcription:

uv run stt models --backend hf --download whisper-my-small
uv run stt transcribe recording.mp3 \
  --backend hf --model whisper-my-small

Use --backend BACKEND --download MODEL for each additional model needed. The command does not prefetch the other cards in that backend.

Implemented backends

Backend	Runtime	auto device behavior	Implemented scope
omniasr-torch	Meta PyTorch/fairseq2	CUDA, then MPS, then CPU; a failed MPS decode is retried on CPU	Upstream omniASR v2 CTC, LLM, and Unlimited cards
omniasr-gguf	CrispASR/ggml	CrispASR chooses the native backend; Metal is available on Apple Silicon	Selected downloadable GGUF conversions up to 1B
hf	Hugging Face Transformers	CUDA, then MPS, then CPU; no automatic CPU retry	Burmese Whisper fine-tunes, MMS-1B, SeamlessM4T v2, and w2v-BERT
dolphin	DataoceanAI Dolphin	CUDA when available, otherwise CPU; MPS must be requested explicitly	Public base and small checkpoints

The omniasr-torch MPS path is a repository policy: the adapter passes device="mps" to the PyTorch pipeline and catches decode failures so it can reload on CPU. It is not a claim that fairseq2 officially supports Metal or that every fairseq2 operation is implemented by MPS. A result recorded after retry must identify CPU as the resolved device.

Model behavior

The fairseq2 adapter accepts only the v2 cards listed by the CLI. CTC and non-Unlimited LLM cards reject audio longer than 40 seconds. Unlimited cards are the upstream long-audio variants.

The omniASR models are trained to emit spoken-form transcripts without punctuation or capitalization. The adapter does not add either. This behavior is documented in Meta's inference guide.

For SeamlessM4T v2, the upstream language table lists Burmese (mya, Mymr) as source speech and source text, with target text only. This repository uses its speech-to-text path; Burmese target speech is not supported by that card.

Dolphin's upstream SPEECH_LENGTH = 30 controls when its CLI chooses long-form transcription and caps VAD segments at 30 seconds. It is not evidence that the single-file feature extractor is a fixed Whisper 30-second window. This adapter windows input at 30 seconds as its own integration policy.

Defaults and precision

The current defaults are:

Backend	Default model
omniasr-torch	omniASR_LLM_Unlimited_7B_v2
omniasr-gguf	llm-unlimited-300m-v2
hf	seamless-m4t-v2
dolphin	small

Defaults do not cause setup to download weights. A transcription or explicit model-download command fetches only the selected model.

Dolphin, GGUF, and Transformers downloads are pinned to immutable upstream revisions. A plain stt compare and --all-cached-models run Hugging Face in cached-files-only mode, so a missing file cannot trigger a download after its cache check. An explicit --only hf comparison or transcription retains normal first-use download behavior.

Dolphin downloads are restricted to the five files its loader consumes at a pinned Hugging Face revision. The adapter verifies every file, including train.yaml, against its SHA-256 digest before passing the cache to Dolphin. GGUF URLs are likewise pinned to immutable Hugging Face revisions, and the adapter verifies the selected model's SHA-256 digest before native loading. The fairseq2 adapter verifies each returned checkpoint's exact byte length and the shared tokenizer's byte length and SHA-256 digest. Upstream does not publish checkpoint digests, so its checkpoint check detects incomplete or wrong-sized files but cannot detect a same-sized modification. A failed check leaves the cache untouched for inspection or quarantine.

For omniasr-torch, --dtype auto measures float16 and bfloat16 on an accelerator. On CPU it uses float32 when the machine has the adapter's estimated memory headroom and bfloat16 otherwise. The Transformers backend currently uses float32 by default. GGUF precision is part of the selected file.

Upstream identity and licences

• Meta Omnilingual ASR publishes the model table and Apache-2.0 code and weights. The exact v2 asset definitions are in the upstream asset cards.
• The GGUF files come from the exact cstr/*-GGUF URLs encoded in the adapter. Check both the conversion card and the original omniASR terms before redistribution.
• The implemented Burmese Whisper large card is Apache-2.0; the medium, small, and stock Whisper cards retain their own card metadata.
• MMS-1B and SeamlessM4T v2 are CC-BY-NC-4.0.
• The implemented Burmese w2v-BERT fine-tune is MIT.
• Dolphin publishes its code and checkpoints under Apache-2.0.

Model licences are separate from this harness's MIT licence. Recheck the exact upstream card and revision before redistribution or product use.