OpenVoiceOS STT Plugin - ONNX ASR

September 1, 2026 · View on GitHub

An OpenVoiceOS Speech-to-Text plugin backed by the lightweight onnx-asr library. This plugin runs offline and supports high-performance models like Nvidia Canary, Parakeet, and OpenAI Whisper via ONNX Runtime.

Description

This plugin enables OpenVoiceOS to use state-of-the-art ASR models exported to ONNX. It leverages the onnx-asr package which provides a unified interface for running various architectures (NeMo, Whisper, GigaAM, etc.) without heavy dependencies like PyTorch.

Install

To install the plugin, use pip. You also need to ensure the backend dependencies are installed.

pip install ovos-stt-plugin-onnx-asr

To run this plugin as a standalone server in a container, see docs/docker.md and docs/deployment.md. For which model families load and how, see docs/models.md.

Configuration

Configure the plugin in your mycroft.conf or user config.

{
  "stt": {
    "module": "ovos-stt-plugin-onnx-asr",
    "ovos-stt-plugin-onnx-asr": {
      "model": "nemo-canary-1b-v2",
      "quantization": "int8"
    }
  }
}

Per-language model resolution

The plugin ships a built-in best-model-per-language registry (defaults.LANG_DEFAULTS, ~90 languages — dedicated fine-tunes from the OpenVoiceOS/stt-asr-onnx collection, with parakeet-tdt-0.6b-v3 and whisper-base as multilingual coverage). Whisper ONNX exports are supported like any other model. The model for a request's language resolves in this order, trying the full BCP-47 tag before the nearest usable tag at each level:

lang2model in the plugin config
ONNX_ASR_DEFAULT_<LANG> environment variables — underscores map to dashes, so ONNX_ASR_DEFAULT_PT=... sets pt and ONNX_ASR_DEFAULT_PT_BR=... sets pt-BR (handy for containers)
the configured model
the built-in registry
nemo-canary-1b-v2, for a language the registry does not hold

Language tags are compared as OVOS-INTENT-2 §2 specifies, via ovos-spec-tools: case insensitively, with underscores accepted for hyphens, falling back to the nearest usable tag. So en_US, en-GB and EN all reach the en entry, while ga (Irish) never reaches the gaa (Ga) one.

A configured model serves every language, because it is your instruction and a registry entry is only the best guess for a language you said nothing about. To let the registry pick per language, leave model unset and name exceptions in lang2model.

Models load lazily on the first request for their language and stay cached in memory, so a single instance (or one ovos-stt-server container) serves every language with the best available model.

A registry model whose repository holds no quantized weights loads fp32, even with quantization set, so that setting cannot take a language away.

Configuration Options

Option	Default	Description
`model`	`nemo-canary-1b-v2`	The model ID to load. Can be a specific alias (like `nemo-parakeet-tdt-0.6b-v3`) or a Hugging Face repo ID. Set it and it serves every language, ahead of the built-in registry; leave it unset for per-language routing.
`lang2model`	`{}`	Optional per-language routing map, e.g. `{"ru": "gigaam-v2-rnnt", "gl": "OpenVoiceOS/proxectonos-gl-conformer-ctc-large-onnx"}`. The model for a language loads lazily on the first request and stays cached in memory, so a single instance (or a single `ovos-stt-server` container) can serve every configured language with the best model for each. Unmapped languages fall back to `model`.
`quantization`	`null`	Set to `"int8"` to load the quantized weights for faster, lower-memory CPU inference. Requires the repo to ship `*.int8.onnx` files; loading fails if they are absent, except for a model the built-in registry picked, which then loads fp32. int8 trades a small accuracy drop (typically a few WER points, less on larger models) for ~3-4x smaller models.
`use_cuda`	`false`	Run on the GPU via the CUDA execution provider (with a CPU fallback).
`providers`	`null`	Explicit list of onnxruntime execution providers, e.g. `["CUDAExecutionProvider", "CPUExecutionProvider"]` or `["TensorrtExecutionProvider"]`. Takes precedence over `use_cuda`.
`cpu_models_only`	`false`	Restrict model selection to models practical on CPU-only hardware. See below.

CPU-only deployments

A satellite or server that will only ever run on CPU should set cpu_models_only rather than trust every operator, or every future config change, to keep clear of a model that expects a GPU. On, it drops any model whose id advertises a parameter count of 0.6B or more (the size the catalogue already writes into model ids like nemo-canary-1b-v2 or qwen3-asr-0.6b-onnx) from the built-in per-language registry, and it swaps the ultimate fallback from nemo-canary-1b-v2 (1B) to whisper-base (74M, the plugin's other multilingual coverage model).

{
  "stt": {
    "module": "ovos-stt-plugin-onnx-asr",
    "ovos-stt-plugin-onnx-asr": {
      "cpu_models_only": true
    }
  }
}

Naming an excluded model explicitly, through model or lang2model, is refused at startup rather than silently swapped for a different one: the config is wrong, and running a different model than the one named is worse than saying so. cpu_models_only is off by default, so leaving it unset keeps every model in the catalogue selectable, exactly as before the option existed.

GPU acceleration

To run on the GPU, install onnxruntime-gpu (in place of the default onnxruntime) with a matching CUDA/cuDNN runtime, then set use_cuda:

{
  "stt": {
    "module": "ovos-stt-plugin-onnx-asr",
    "ovos-stt-plugin-onnx-asr": {
      "model": "nemo-parakeet-tdt-0.6b-v3",
      "use_cuda": true
    }
  }
}

For finer control (e.g. TensorRT) set providers directly; it overrides use_cuda.

Supported Models

The model option accepts either a built-in onnx-asr alias or any Hugging Face repo id whose config.json declares a supported model_type (NeMo Conformer/FastConformer with CTC, RNN-T, TDT or Canary/AED decoder, Whisper, Vosk, GigaAM, T-one or wav2vec2-CTC — plus, with the fork pin below: speech-llm, espnet-ctc/espnet-aed, granite-nar, sensevoice, moonshine and wav2vec2-adapters). Streaming and TTS checkpoints are not supported.

language is meaningful for Whisper, Canary/AED, speech-llm and wav2vec2-adapters models (target_language for Canary); the plugin passes it automatically only to those families.

Choosing a model: architecture tiers

Models in the STT/ASR - onnx collection span several architecture families. They trade accuracy against size and speed in a fairly consistent order. RTF = real-time factor on a mid-range desktop CPU (AMD Ryzen 5 7600); lower is faster, below 1.0 is faster than real time.

Tier	Families	Size	CPU speed	When to use
Speech-LLM	`speech-llm` (Qwen3-ASR, Canary-Qwen, Granite, Voxtral, AMALIA), `granite-nar`	1–19 GB	RTF ~0.3–1.3 (int8); NAR variant ~0.3 fp32	Best accuracy available. Use on a server with RAM to spare, or when transcription quality matters more than latency. The 9B models effectively need a GPU.
Attention encoder-decoder	`nemo-conformer-aed` (Canary, Cohere Transcribe), Whisper large / `espnet-aed`	1–8 GB	RTF ~0.2–0.9 (int8)	Strong accuracy with punctuation and casing. Good server default. Cohere Transcribe covers 14 languages at RTF ~0.19 int8.
Conformer CTC / transducer	`nemo-conformer-ctc/rnnt/tdt` (Parakeet), GigaAM, Vosk	0.1–2.5 GB	RTF ~0.05–0.3	The practical sweet spot for assistants: fast, small, accurate for their languages. `nemo-parakeet-tdt-0.6b-v3` is the plugin default for good reason.
Compact Whisper	whisper small/medium, distil, lite-whisper	0.2–1 GB (int8)	RTF ~0.3–1.0	Many language fine-tunes only exist as Whisper checkpoints. Use the fine-tune for your language when one exists.
Tiny specialists	`sensevoice` (zh/en/ja/ko/yue), `moonshine` (en)	60–950 MB	RTF ~0.02–0.04	Fastest options by far. Use on constrained hardware (Raspberry Pi class) for their languages.
wav2vec2 CTC / MMS	`wav2vec2-ctc`, `wav2vec2-adapters` (MMS, 1100+ languages)	~1 GB int8 base	RTF ~0.15	Lowest accuracy tier: character-level output, no punctuation or casing, narrow training domains. But for hundreds of languages this is the only model that exists. Something is better than nothing.

Rules of thumb:

A dedicated fine-tune for your language beats a bigger multilingual model more often than not. Check the collection for your language tag first.
If your language has a Whisper or Conformer fine-tune, prefer it over MMS. Use wav2vec2-adapters (MMS) when nothing else covers the language.
int8 quantization roughly quarters the size. For autoregressive models it is usually ~2x faster on CPU; for single-pass models (NAR, CTC) it can be slower — check the model card, each states its measured numbers.
Speech-LLM and AED models emit punctuation and casing; CTC-family models generally do not.

Built-in aliases (Nvidia NeMo, Whisper, GigaAM, Vosk, T-one)

Alias	Language(s)
`nemo-canary-1b-v2`	Multilingual (en, de, fr, es, …)
`nemo-parakeet-tdt-0.6b-v3`	Multilingual (25 European langs)
`nemo-parakeet-tdt-0.6b-v2` / `nemo-parakeet-ctc-0.6b` / `nemo-parakeet-rnnt-0.6b`	English
`gigaam-v3-ctc` / `gigaam-v3-rnnt` / `gigaam-v2-ctc` / `gigaam-v2-rnnt`	Russian
`nemo-fastconformer-ru-ctc` / `nemo-fastconformer-ru-rnnt`	Russian
`alphacep/vosk-model-ru` / `alphacep/vosk-model-small-ru` / `t-tech/t-one`	Russian
`whisper-base` / `onnx-community/whisper-large-v3-turbo`	Multilingual

OpenVoiceOS curated models

Curated single-language and regional models live in the OpenVoiceOS/stt-asr-onnx collection, converted from reputable NeMo Conformer/Parakeet checkpoints and loaded by repo id. Repos are named <author>-<model>-onnx to avoid collisions between same-named finetunes. Most ship both fp32 and int8 weights, so quantization: "int8" works (a few large models are fp32-only; NeMo CTC exports like the Citrinet and FastConformer-Hybrid pc fleets ship fp32 only).

Every model in the collection, grouped by language:

Language	`model`	Architecture
Indian languages (multilingual)	`OpenVoiceOS/artpark-iisc-vaani-fastconformer-multi-onnx`	FastConformer
Arabic	`OpenVoiceOS/stt_ar_fastconformer_hybrid_large_pc_v1.0_onnx`	FastConformer-Hybrid (CTC, punct+case)
Arabic	`OpenVoiceOS/stt_ar_fastconformer_hybrid_large_pcd_v1.0_onnx`	FastConformer-Hybrid (CTC, punct+case)
Armenian	`OpenVoiceOS/stt_hy-AM_citrinet_512_armenian-CV17.0_onnx`	Citrinet CTC
Armenian	`OpenVoiceOS/stt_hy_fastconformer_hybrid_large_pc_onnx`	FastConformer-Hybrid (CTC, punct+case)
Assamese	`OpenVoiceOS/ai4bharat-indicconformer-as-onnx`	IndicConformer CTC
Bangla	`OpenVoiceOS/ai4bharat-indicconformer-bn-onnx`	IndicConformer CTC
Basque	`OpenVoiceOS/hitz-eu-conformer-transducer-large-v2-onnx`	Conformer RNN-T
Basque	`OpenVoiceOS/stt-eu-conformer-ctc-large-onnx`	Conformer CTC
Basque	`OpenVoiceOS/stt-eu-conformer-transducer-large-onnx`	Conformer RNN-T
Belarusian	`OpenVoiceOS/nvidia-be-conformer-ctc-large-onnx`	Conformer CTC
Belarusian	`OpenVoiceOS/nvidia-be-conformer-transducer-large-onnx`	Conformer RNN-T
Belarusian	`OpenVoiceOS/stt_be_fastconformer_hybrid_large_pc_onnx`	FastConformer-Hybrid (CTC, punct+case)
Bodo	`OpenVoiceOS/ai4bharat-indicconformer-brx-onnx`	IndicConformer CTC
Catalan	`OpenVoiceOS/neongeckocom-stt_ca_citrinet_512_gamma_0_25_onnx`	Citrinet CTC
Catalan	`OpenVoiceOS/nvidia-ca-conformer-ctc-large-onnx`	Conformer CTC
Catalan	`OpenVoiceOS/nvidia-ca-conformer-transducer-large-onnx`	Conformer RNN-T
Catalan	`OpenVoiceOS/whisper-large-v3-ca-punctuated-3370h-onnx`	Whisper
Catalan / Spanish	`OpenVoiceOS/stt-ca-es-conformer-transducer-large-onnx`	Conformer RNN-T
Catalan / Spanish	`OpenVoiceOS/whisper-large-v3-tiny-caesar-onnx`	Whisper
Chinese	`OpenVoiceOS/stt_zh_citrinet_1024_gamma_0_25_onnx`	Citrinet CTC
Chinese	`OpenVoiceOS/stt_zh_citrinet_512_onnx`	Citrinet CTC
Croatian	`OpenVoiceOS/nvidia-hr-conformer-ctc-large-onnx`	Conformer CTC
Croatian	`OpenVoiceOS/nvidia-hr-conformer-transducer-large-onnx`	Conformer RNN-T
Croatian	`OpenVoiceOS/stt_hr_fastconformer_hybrid_large_pc_onnx`	FastConformer-Hybrid (CTC, punct+case)
Danish	`OpenVoiceOS/nvidia-parakeet-rnnt-110m-da-dk-onnx`	Parakeet RNN-T
Dogri	`OpenVoiceOS/ai4bharat-indicconformer-doi-onnx`	IndicConformer CTC
Dutch	`OpenVoiceOS/neongeckocom-stt_nl_citrinet_512_gamma_0_25_onnx`	Citrinet CTC
Dutch	`OpenVoiceOS/stt_nl_fastconformer_hybrid_large_pc_onnx`	FastConformer-Hybrid (CTC, punct+case)
Dutch	`OpenVoiceOS/yuriyvnv-parakeet-tdt-0.6b-nl-onnx`	Parakeet TDT
English	`OpenVoiceOS/neongeckocom-stt_en_citrinet_512_gamma_0_25_onnx`	Citrinet CTC
English	`OpenVoiceOS/nvidia-en-conformer-ctc-large-onnx`	Conformer CTC
English	`OpenVoiceOS/nvidia-en-conformer-ctc-small-onnx`	Conformer CTC
English	`OpenVoiceOS/nvidia-en-conformer-transducer-large-onnx`	Conformer RNN-T
English	`OpenVoiceOS/nvidia-en-conformer-transducer-xlarge-onnx`	Conformer RNN-T
English	`OpenVoiceOS/nvidia-parakeet-ctc-1.1b-onnx`	Parakeet CTC
English	`OpenVoiceOS/nvidia-parakeet-rnnt-1.1b-onnx`	Parakeet RNN-T
English	`OpenVoiceOS/nvidia-parakeet-tdt-1.1b-onnx`	Parakeet TDT
English	`OpenVoiceOS/nvidia-parakeet-tdt_ctc-110m-onnx`	Parakeet TDT+CTC
English	`OpenVoiceOS/stt_en_citrinet_1024_gamma_0_25_onnx`	Citrinet CTC
English	`OpenVoiceOS/stt_en_citrinet_256_gamma_0_25_onnx`	Citrinet CTC
English	`OpenVoiceOS/stt_en_citrinet_512_gamma_0_25_onnx`	Citrinet CTC
English	`OpenVoiceOS/stt_en_fastconformer_hybrid_large_pc_onnx`	FastConformer-Hybrid (CTC, punct+case)
Esperanto	`OpenVoiceOS/nvidia-eo-conformer-ctc-large-onnx`	Conformer CTC
Esperanto	`OpenVoiceOS/nvidia-eo-conformer-transducer-large-onnx`	Conformer RNN-T
Estonian	`OpenVoiceOS/yuriyvnv-parakeet-tdt-0.6b-et-onnx`	Parakeet TDT
Fanti	`OpenVoiceOS/misterkissi-w2v2-lg-xls-r-300m-fante-onnx`	wav2vec2 CTC
Filipino	`OpenVoiceOS/stt-tl-fastconformer-hybrid-large-onnx`	FastConformer-Hybrid
Fon	`OpenVoiceOS/misterkissi-whisper-small-fongbe-onnx`	Whisper
French	`OpenVoiceOS/neongeckocom-stt_fr_citrinet_512_gamma_0_25_onnx`	Citrinet CTC
French	`OpenVoiceOS/nvidia-fr-conformer-ctc-large-onnx`	Conformer CTC
French	`OpenVoiceOS/nvidia-fr-conformer-transducer-large-onnx`	Conformer RNN-T
French	`OpenVoiceOS/stt_fr_citrinet_1024_gamma_0_25_onnx`	Citrinet CTC
French	`OpenVoiceOS/stt_fr_fastconformer_hybrid_large_pc_onnx`	FastConformer-Hybrid (CTC, punct+case)
Galician	`OpenVoiceOS/proxectonos-gl-conformer-ctc-large-onnx`	Conformer CTC
Georgian	`OpenVoiceOS/stt_ka_fastconformer_hybrid_large_pc_onnx`	FastConformer-Hybrid (CTC, punct+case)
German	`OpenVoiceOS/neongeckocom-stt_de_citrinet_512_gamma_0_25_onnx`	Citrinet CTC
German	`OpenVoiceOS/nvidia-de-conformer-ctc-large-onnx`	Conformer CTC
German	`OpenVoiceOS/nvidia-de-conformer-transducer-large-onnx`	Conformer RNN-T
German	`OpenVoiceOS/stt_de_citrinet_1024_onnx`	Citrinet CTC
German	`OpenVoiceOS/stt_de_fastconformer_hybrid_large_pc_onnx`	FastConformer-Hybrid (CTC, punct+case)
Gujarati	`OpenVoiceOS/ai4bharat-indicconformer-gu-onnx`	IndicConformer CTC
Haitian Creole	`OpenVoiceOS/misterkissi-whisper-small-haitian-creole-onnx`	Whisper
Hindi	`OpenVoiceOS/ai4bharat-indicconformer-hi-onnx`	IndicConformer CTC
Hindi	`OpenVoiceOS/artpark-iisc-vaani-fastconformer-hi-onnx`	FastConformer
Ikposo	`OpenVoiceOS/misterkissi-w2v2-lg-xls-r-300m-kposo-onnx`	wav2vec2 CTC
Ikposo	`OpenVoiceOS/misterkissi-whisper-small-kposo-onnx`	Whisper
Ga	`OpenVoiceOS/misterkissi-w2v2-lg-xls-r-300m-ga-onnx`	wav2vec2 CTC — Ga, the Kwa language of Accra (`gaa`), not Irish
Italian	`OpenVoiceOS/neongeckocom-stt_it_citrinet_512_gamma_0_25_onnx`	Citrinet CTC
Italian	`OpenVoiceOS/nvidia-it-conformer-ctc-large-onnx`	Conformer CTC
Italian	`OpenVoiceOS/nvidia-it-conformer-transducer-large-onnx`	Conformer RNN-T
Italian	`OpenVoiceOS/stt_it_fastconformer_hybrid_large_pc_onnx`	FastConformer-Hybrid (CTC, punct+case)
Japanese	`OpenVoiceOS/nvidia-parakeet-tdt_ctc-0.6b-ja-onnx`	Parakeet TDT+CTC
Kabyle	`OpenVoiceOS/misterkissi-w2v2-lg-xls-r-1b-kabyle-onnx`	wav2vec2 CTC
Kabyle	`OpenVoiceOS/nvidia-kab-conformer-transducer-large-onnx`	Conformer RNN-T
Kannada	`OpenVoiceOS/ai4bharat-indicconformer-kn-onnx`	IndicConformer CTC
Kannada	`OpenVoiceOS/artpark-iisc-vaani-fastconformer-kn-onnx`	FastConformer
Kasem	`OpenVoiceOS/misterkissi-w2v2-lg-xls-r-300m-kasem-onnx`	wav2vec2 CTC
Kashmiri	`OpenVoiceOS/ai4bharat-indicconformer-ks-onnx`	IndicConformer CTC
Kazakh / Russian	`OpenVoiceOS/stt_kk_ru_fastconformer_hybrid_large_onnx`	FastConformer-Hybrid (CTC, punct+case)
Kinyarwanda	`OpenVoiceOS/nvidia-rw-conformer-ctc-large-onnx`	Conformer CTC
Kinyarwanda	`OpenVoiceOS/nvidia-rw-conformer-transducer-large-onnx`	Conformer RNN-T
Konkani	`OpenVoiceOS/ai4bharat-indicconformer-kok-onnx`	IndicConformer CTC
Korean	`OpenVoiceOS/stt_kr_citrinet1024_PublicCallCenter_1000H_onnx`	Citrinet CTC
Lingala	`OpenVoiceOS/misterkissi-w2v2-lg-xls-r-300m-lingala-onnx`	wav2vec2 CTC
Lomwe	`OpenVoiceOS/misterkissi-whisper-small-lomwe-onnx`	Whisper
Maithili	`OpenVoiceOS/ai4bharat-indicconformer-mai-onnx`	IndicConformer CTC
Malagasy	`OpenVoiceOS/misterkissi-whisper-small-malagasy-onnx`	Whisper
Malayalam	`OpenVoiceOS/ai4bharat-indicconformer-ml-onnx`	IndicConformer CTC
Malayalam	`OpenVoiceOS/artpark-iisc-vaani-fastconformer-ml-onnx`	FastConformer
Manipuri	`OpenVoiceOS/ai4bharat-indicconformer-mni-onnx`	IndicConformer CTC
Marathi	`OpenVoiceOS/ai4bharat-indicconformer-mr-onnx`	IndicConformer CTC
Nepali	`OpenVoiceOS/ai4bharat-indicconformer-ne-onnx`	IndicConformer CTC
Nyankole	`OpenVoiceOS/misterkissi-w2v2-lg-xls-r-300m-runyankore-onnx`	wav2vec2 CTC
Odia	`OpenVoiceOS/ai4bharat-indicconformer-or-onnx`	IndicConformer CTC
Odia	`OpenVoiceOS/artpark-iisc-vaani-fastconformer-or-onnx`	FastConformer
Persian	`OpenVoiceOS/nvidia-fa-fastconformer-hybrid-large-onnx`	FastConformer-Hybrid
Persian	`OpenVoiceOS/stt_fa_fastconformer_hybrid_large_onnx`	FastConformer-Hybrid (CTC, punct+case)
Polish	`OpenVoiceOS/stt_pl_fastconformer_hybrid_large_pc_onnx`	FastConformer-Hybrid (CTC, punct+case)
Polish	`OpenVoiceOS/yuriyvnv-parakeet-tdt-0.6b-pl-onnx`	Parakeet TDT
Portuguese	`OpenVoiceOS/neongeckocom-stt_pt_citrinet_512_gamma_0_25_onnx`	Citrinet CTC
Portuguese	`OpenVoiceOS/stt_pt_fastconformer_hybrid_large_pc_onnx`	FastConformer-Hybrid (CTC, punct+case)
Portuguese	`OpenVoiceOS/whisper-large-v3-pt-onnx`	Whisper
Portuguese	`OpenVoiceOS/whisper-medium-pt-onnx`	Whisper
Portuguese	`OpenVoiceOS/whisper-small-pt-onnx`	Whisper
Portuguese	`OpenVoiceOS/yuriyvnv-parakeet-tdt-0.6b-pt-onnx`	Parakeet TDT
Punjabi	`OpenVoiceOS/ai4bharat-indicconformer-pa-onnx`	IndicConformer CTC
Russian	`OpenVoiceOS/nvidia-ru-conformer-ctc-large-onnx`	Conformer CTC
Russian	`OpenVoiceOS/nvidia-ru-conformer-transducer-large-onnx`	Conformer RNN-T
Russian	`alphacep/vosk-model-ru`	Vosk
Sanskrit	`OpenVoiceOS/ai4bharat-indicconformer-sa-onnx`	IndicConformer CTC
Santali	`OpenVoiceOS/ai4bharat-indicconformer-sat-onnx`	IndicConformer CTC
Shona	`OpenVoiceOS/misterkissi-whisper-small-shona-onnx`	Whisper
Sindhi	`OpenVoiceOS/ai4bharat-indicconformer-sd-onnx`	IndicConformer CTC
Slovenian	`OpenVoiceOS/yuriyvnv-parakeet-tdt-0.6b-sl-onnx`	Parakeet TDT
Southern Sotho	`OpenVoiceOS/misterkissi-whisper-small-sesotho-onnx`	Whisper
Spanish	`OpenVoiceOS/bsc-lt-los-conformer-transducer-large-onnx`	Conformer RNN-T
Spanish	`OpenVoiceOS/bsc-lt-los-conformer-transducer-large-punctuated-onnx`	Conformer RNN-T
Spanish	`OpenVoiceOS/hitz-bbs-s2tc-conformer-transducer-large-onnx`	Conformer RNN-T
Spanish	`OpenVoiceOS/hitz-eseu-conformer-transducer-large-onnx`	Conformer RNN-T
Spanish	`OpenVoiceOS/neongeckocom-stt_es_citrinet_512_gamma_0_25_onnx`	Citrinet CTC
Spanish	`OpenVoiceOS/nvidia-es-conformer-ctc-large-onnx`	Conformer CTC
Spanish	`OpenVoiceOS/nvidia-es-conformer-transducer-large-onnx`	Conformer RNN-T
Spanish	`OpenVoiceOS/parakeet-rnnt-1.1b-cv17-es-ep18-1270h-onnx`	Parakeet RNN-T
Spanish	`OpenVoiceOS/stt_es_citrinet_1024_gamma_0_25_onnx`	Citrinet CTC
Spanish	`OpenVoiceOS/stt_es_citrinet_512_onnx`	Citrinet CTC
Spanish	`OpenVoiceOS/stt_es_fastconformer_hybrid_large_pc_onnx`	FastConformer-Hybrid (CTC, punct+case)
Tajik	`alphacep/vosk-model-tg`	Vosk
Tamil	`OpenVoiceOS/ai4bharat-indicconformer-ta-onnx`	IndicConformer CTC
Telugu	`OpenVoiceOS/ai4bharat-indicconformer-te-onnx`	IndicConformer CTC
Telugu	`OpenVoiceOS/artpark-iisc-vaani-fastconformer-te-onnx`	FastConformer
Tigre	`OpenVoiceOS/misterkissi-whisper-small-tigre-onnx`	Whisper
Tsonga	`OpenVoiceOS/misterkissi-w2v2-lg-xls-r-300m-tsonga-onnx`	wav2vec2 CTC
Tswana	`OpenVoiceOS/misterkissi-whisper-small-setswana-onnx`	Whisper
Ukrainian	`OpenVoiceOS/neongeckocom-stt_uk_citrinet_512_gamma_0_25_onnx`	Citrinet CTC
Ukrainian	`OpenVoiceOS/stt_uk_citrinet_1024_gamma_0_25_onnx`	Citrinet CTC
Umbundu	`OpenVoiceOS/misterkissi-whisper-small-umbundu-onnx`	Whisper
Unknown language [ua]	`OpenVoiceOS/stt_ua_fastconformer_hybrid_large_pc_onnx`	FastConformer-Hybrid (CTC, punct+case)
Urdu	`OpenVoiceOS/ai4bharat-indicconformer-ur-onnx`	IndicConformer CTC
Uzbek	`OpenVoiceOS/asr-uz-fastconformer-large-onnx`	FastConformer
Uzbek	`OpenVoiceOS/stt_uz_fastconformer_hybrid_large_pc_onnx`	FastConformer-Hybrid (CTC, punct+case)
Vagla	`OpenVoiceOS/misterkissi-w2v2-lg-xls-r-300m-vagla-onnx`	wav2vec2 CTC
Vai	`OpenVoiceOS/misterkissi-w2v2-lg-xls-r-300m-vai-onnx`	wav2vec2 CTC
Vai	`OpenVoiceOS/misterkissi-whisper-small-vai-onnx`	Whisper
Vietnamese	`OpenVoiceOS/nvidia-parakeet-ctc-0.6b-vietnamese-onnx`	Parakeet CTC
Xhosa	`OpenVoiceOS/misterkissi-w2v2-lg-xls-r-300m-xhosa-onnx`	wav2vec2 CTC
Zulu	`OpenVoiceOS/misterkissi-w2v2-lg-xls-r-300m-zulu-onnx`	wav2vec2 CTC

Every curated repo carries a model card with base_model metadata pointing to the upstream checkpoint and the training source. See the collection for the exhaustive, up-to-date list.

For the full list of built-in aliases and benchmarks, see the onnx-asr repository.

wav2vec2 models

wav2vec2 and XLS-R CTC fine-tunes load by repo id with no extra setup. See docs/models.md.

Fork model families (draft, pinned to a fork branch)

istupakov/onnx-asr does not yet support three model families used by some of our HF exports: ESPnet E-Branchformer (CTC and attention-decoder variants) and Speech-LLM (audio encoder + projector + causal LM, covers both plain Qwen3-ASR-style prompting and the SALM variant used by Canary-Qwen). Until that lands upstream, this plugin can only load those repos against TigreGotico/onnx-asr's integration branch — see requirements.txt. This is why the PR that added this section stays a draft: a git ref in install_requires breaks PyPI publishing, so it must revert to a normal version floor once the upstream PRs merge and release.

Upstream tracking PRs (parity benchmarks and RTF numbers in each):

istupakov/onnx-asr#3 — speech-llm model family (Qwen3-ASR)
istupakov/onnx-asr#4 — ESPnet E-Branchformer models (espnet-ctc / espnet-aed) + w2v-BERT 2.0 preprocessor
istupakov/onnx-asr#5 — SALM encoder shape for speech-llm (Canary-Qwen-2.5B)

No plugin config changes are needed to use these models: set model to the HF repo id as usual, the model type comes from the repo's config.json. The plugin also passes a language hint to Speech-LLM models the same way it does for Whisper and Canary.

Model	Family	HF repo	Fork branch
INESC-ID e-branchformer (European Portuguese), CTC	`espnet-ctc`	OpenVoiceOS/inesc-id-ebranch-w2vbert2-ep-ctc-onnx	`integration`
INESC-ID e-branchformer (European Portuguese), attention decoder	`espnet-aed`	OpenVoiceOS/inesc-id-ebranch-w2vbert2-ep-aed-onnx	`integration`
Qwen3-ASR 0.6B	`speech-llm`	OpenVoiceOS/qwen3-asr-0.6b-onnx	`integration`
Qwen3-ASR 1.7B	`speech-llm`	OpenVoiceOS/qwen3-asr-1.7b-onnx	`integration`
Canary-Qwen 2.5B (SALM)	`speech-llm`	OpenVoiceOS/canary-qwen-2.5b-onnx	`integration` (needs the SALM commit, included)
Granite Speech 3.3 2B	`speech-llm`	OpenVoiceOS/granite-speech-3.3-2b-onnx	`integration`
INESC-ID Whisper large-v3 (European Portuguese fine-tune)	Whisper	OpenVoiceOS/inesc-id-whisperlv3-ft-ep-onnx	none — Whisper is already supported
Camões Whisper (Portuguese)	Whisper	OpenVoiceOS/camoes-whisper-asr-onnx	none — Whisper is already supported

Credits

istupakov/onnx-asr - The underlying library doing the heavy lifting.

This project was funded through the NGI0 Commons Fund, a fund established by NLnet with financial support from the European Commission's Next Generation Internet programme, under the aegis of DG Communications Networks, Content and Technology under grant agreement No 101135429.