Skip to content
OpenASR

Models

The public model catalog lives at model-registry/catalog.json in the open core repository and is signed and verified before the CLI trusts it. The Models page renders a static projection of that catalog — the current source of truth for exact quants, sizes, checksums, and licenses.

Families in the catalog today

Section titled “Families in the catalog today”

The signed catalog grows over time. Treat the Models page as the live list of public packs, sizes, quants, and licenses. Families currently include:

  • Whisper — English-only and multilingual variants across tiny/base/small/medium/large-v3/large-v3-turbo.
  • Qwen3-ASR — multilingual, available in 0.6B and 1.7B sizes.
  • Parakeet-TDT — NVIDIA’s multilingual ASR (parakeet-tdt-0.6b-v3), covering 25 European languages.
  • FireRed — FireRedTeam bilingual ASR (firered-aed-l-v2 and the larger firered2-llm).
  • MiMo-ASR — Xiaomi MiMo end-to-end ASR (mimo-v2.5-asr) for Mandarin, English, and Cantonese.
  • Fun-ASR-Nano — compact Mandarin + English ASR from FunAudioLLM.
  • Granite Speech — IBM Granite multilingual ASR (granite-speech-4.1-2b).
  • MOSS-Transcribe-Diarize — joint transcription and speaker diarization in one pack.
  • SenseVoice — multilingual ASR from FunAudioLLM (sensevoice-small), covering Chinese, Cantonese, English, Japanese, and Korean.
  • Moonshine — compact English ASR.
  • Cohere Transcribe — multilingual ASR.
  • X-ASR — a zh-en bilingual model.
  • Dolphin — a multilingual ASR family: general models (dolphin-base, dolphin-small) covering roughly 40 languages across South/Southeast/Central Asia plus Chinese, and dedicated Chinese-dialect models (dolphin-cn-dialect-base, dolphin-cn-dialect-small) for regional varieties.
  • Hy-MT2 — a speech translation model (not plain transcription).
  • Capability packs — diarization support (speaker segmentation and voice-match embedding models) and FireRed punctuation (Chinese punctuation restoration, firered-punc), pulled automatically when the CLI flags that need them (for example --diarize) are used.

Most ASR entries ship several quantizations (for example fp16, q8_0, q4_k); some packs (the translation and diarization capability packs) ship a single tier. The catalog marks one recommended_quant per model. Don’t hardcode a quant list here — it changes as new quants are published. Check openasr search or the Models page for the live set.

The .oasr package format

Section titled “The .oasr package format”

Every model ships as a single .oasr file — GGUF-backed internally, magic-led with GGUF, and the only user-facing pack format; bare .gguf is not accepted as run input or importer output. openasr verify and openasr show inspect a pack without downloading anything, and native transcription is fail-closed for unsupported or malformed pack inputs.

Installing and using a model

Section titled “Installing and using a model”
bash
openasr search whisper          # find matching catalog entries
openasr pull whisper-small       # download and verify a pack
openasr list                     # see what's installed
openasr transcribe audio.wav --model whisper-small

openasr pull is the explicit, user-initiated download surface. Normal transcription, batch, benchmark, API, and serve flows never auto-download a model outside the CLI’s visible consent prompt (see Current status).

Building your own pack

Section titled “Building your own pack”

openasr model-pack import is a maintainer-only, source-checkout workflow — see the README’s “Building model packs” section if you need it.