STT Plugins¶

In a nutshell

STT stands for Speech-to-Text: this is the part that listens to your spoken words and writes them down as text the assistant can read. It is the same idea as the dictation feature on a phone. Different STT plugins offer different trade-offs in speed, accuracy, and whether they run on your own device or in the cloud, so you can pick the one that suits you. See the Glossary for related terms.

Writing a plugin instead of choosing one? See Writing an STT Plugin.

Formal specification

STT sits inside the audio input service, specified by OVOS-AUDIO-IN-1: Audio Input Service: capture → audio-transformer chain → STT → utterance. The transcript is emitted on ovos.utterance.handle (OVOS-PIPELINE-1 §9.1). See the spec index.

Audio format contract

Everything upstream of an STT plugin, the microphone plugin and any audio transformers, hands over raw PCM in a fixed shape: 16 kHz sample rate, 16-bit samples, mono, little-endian, delivered in 4096-byte chunks by default (the Microphone template's sample_rate/sample_width/sample_channels/chunk_size defaults). A batch STT.execute() plugin gets this bundled into a speech_recognition.AudioData object. A StreamingSTT plugin receives it chunk-by-chunk, still at this same format, unless a deployment explicitly reconfigures the microphone.

Neither the STT nor the StreamingSTT template resamples on the plugin's behalf. If the wrapped model expects a different native rate, converting the incoming 16 kHz PCM is the plugin's own job.

STT (Speech-to-Text) plugins convert spoken audio into text. They are the bridge between the listener and the intent pipeline.

Recommended: onnx-asr

For offline, on-device recognition, ovos-stt-plugin-onnx-asr is the recommended starting point: it runs entirely through ONNX Runtime, with no PyTorch/transformers dependency, and covers NeMo Parakeet/Canary, Whisper and wav2vec2 model families. Cloud STT (e.g. ovos-stt-plugin-azure) is a fair choice when the local device doesn't have the compute budget for on-device recognition.

ovos-stt-plugin-onnx-asr is rated Beta while several alternatives are Stable. It is recommended here for its offline quality and footprint. Per the Maturity Scale, maturity is not the same as a recommendation.

Footprint: as noted in the plugin's own entry, most models in the OpenVoiceOS/stt-asr-onnx collection ship both fp32 and int8 weights. Pick int8 for the lower-RAM, lower-compute option on constrained hardware. Beyond quantization, footprint is mostly driven by model family/size: a small Parakeet/wav2vec2-class model is far lighter than a large Whisper-class one, independent of the plugin wrapping it.

Using an STT plugin¶

Install a plugin and point your mycroft.conf at it:

pip install ovos-stt-plugin-onnx-asr

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

The roster below lists available plugins with a recommended starting configuration for each. This may differ from a plugin's built-in default, noted where relevant.

STT Plugins Reference¶

Code license is the SPDX license of the plugin's own repository. Where the plugin wraps a separately-licensed model, that is called out under "model".

Plugin	Description	License	Maturity
ovos-stt-plugin-wav2vec	OVOS plugin for Wav2Vec2	Apache-2.0 (model: see model card)	Stable
ovos-stt-plugin-azure	Microsoft Azure cloud speech-to-text.	Apache-2.0 (cloud service, separate Microsoft terms)	Stable
ovos-stt-plugin-chromium	Speech-to-text using the Google Chrome browser speech API.	Apache-2.0 (cloud service, unofficial Google endpoint, separate Google terms)	Stable
ovos-stt-plugin-mms	OVOS plugin for The Massively Multilingual Speech (MMS) project. Warning: archived. MMS models also run under ovos-stt-plugin-wav2vec2, installed with `pip install ovos-stt-plugin-wav2vec`, since the PyPI package keeps the old name, or use onnx-asr, which covers the same wav2vec2 families.	Apache-2.0 (model: see model card)	Deprecated
ovos-stt-plugin-server	Companion plugin for an OVOS STT Server. Install name is `ovos-stt-plugin-server`; the repo is `ovos-stt-server-plugin` (words swapped). Sends audio to public community servers unless you set `urls`	Apache-2.0	Stable
ovos-stt-http-server	Turn any OVOS STT plugin into a micro service!	Apache-2.0	Stable
ovos-stt-plugin-whisper	OpenVoiceOS STT plugin for Whisper, using the transformers library	Apache-2.0 (code default model: `base`. Configure a larger model such as openai/whisper-large-v3 for higher accuracy)	Beta
ovos-stt-plugin-whispercpp	OpenVoiceOS STT plugin for whispercpp	Apache-2.0 (model: see model card)	Stable
ovos-stt-plugin-fasterwhisper	OpenVoiceOS STT plugin for Faster Whisper	Apache-2.0 (code default model: `large-v3-turbo`)	Stable
ovos-stt-plugin-nemo	OpenVoiceOS STT plugin for Nemo, GPU is strongly recommended	Apache-2.0 (model: see model card)	Stable
ovos-stt-plugin-whisper-lm	OpenVoiceOS STT plugin for Whisper-LM-transformers, KenLM and Large language model integration with Whisper ASR models implemented in Hugging Face library.	Apache-2.0 (model: see model card)	Stable
ovos-stt-plugin-citrinet	OpenVoiceOS STT plugin. Warning: deprecated by its own README in favor of onnx-asr, which runs the same Citrinet models and more.	Apache-2.0 (model: see model card)	Stable
ovos-stt-plugin-nos	Galician STT using Proxecto Nós wav2vec2 models. Warning: archived, superseded by ovos-stt-plugin-wav2vec2, installed with `pip install ovos-stt-plugin-wav2vec`, since the PyPI package keeps the old name, or use onnx-asr, which covers the same wav2vec2 families.	Apache-2.0 (model: see model card)	Deprecated
ovos-stt-plugin-HiTZ	OpenVoiceOS STT plugin for Basque models trained by HiTZ. Warning: archived, deprecated.	see repo (no license file)	Deprecated
ovos-stt-plugin-vosk	Mycroft STT plugin for Vosk	Apache-2.0 (model: see model card)	Stable
ovos-stt-plugin-onnx-asr	Runs onnx-asr models (NeMo Parakeet/Canary, Whisper, wav2vec2, …) fully offline via ONNX Runtime, a strong default for on-device, offline recognition.	Apache-2.0 (model: see model card)	Beta

Maturity reflects repository health (age, activity, open issues/PRs, in-repo docs), not version. See the Maturity Scale.

License and Maturity are independent axes

The License column reports what the repository itself declares. "No license file" just means no SPDX license was found, not that the code is unmature. The Maturity column reports repository health (age, activity, issues/PRs, docs). A plugin can be Mature and still ship no license file. A plugin can be Stable with a permissive license but thin docs. Don't read one column as implying the other.

Choosing an ASR model family¶

The plugins above wrap a handful of model families with very different trade-offs. Each entry links the architecture paper and a canonical model card, which is where the authoritative detail on training data and per-language accuracy lives.

Whisper (plugins: whisper, whispercpp, fasterwhisper, whisper-lm; also via onnx-asr). The broadest language coverage and the best robustness to accents and noise, because it trains on 680k hours of weakly supervised web audio (Radford et al. 2022). The large-v3 checkpoint adds millions of pseudo-labeled hours (model card). The same weak supervision causes its signature failure: hallucinated phrases on silence, which is why the listener ships a hallucination filter. It is chunk-based, not streaming, and the large checkpoints are the heaviest option here. Prefer large-v3-turbo and int8 quantization on CPU.

NVIDIA NeMo Conformer / Parakeet / Canary (plugins: nemo, citrinet; also via onnx-asr). Best accuracy-per-parameter and much lower latency inside their trained language set (Conformer, Fast Conformer). Parakeet checkpoints are single language and ignore lang (parakeet-tdt-0.6b-v2 card, ~120k hours English). Canary is the multilingual branch (canary-1b-v2 card, ~1.7M hours across 25 European languages). Citrinet is the older CTC-only sibling, kept for specific languages such as the Catalan default.

wav2vec 2.0 / XLS-R / MMS (plugins: wav2vec, mms, nos, HiTZ; also via onnx-asr). Self-supervised pretraining plus a thin fine-tune (Baevski et al. 2020), which is why this family dominates low-resource languages. XLS-R pretrains on ~500k hours across 128 languages, and MMS reaches 1,100+ languages with per-language adapters (mms-1b-all card). Lighter than Whisper, but per-language quality tracks how much fine-tuning data that language got.

Vosk / Kaldi (plugin: vosk). Classical HMM-hybrid decoding with ~50MB per-language models (project page, repo). The lightest footprint and true streaming, at clearly lower accuracy than the neural families. The right choice for constrained hardware and the wrong one for noisy rooms.

Learn more¶

Full technical detail per plugin (repository link, license notes, default configuration) lives on the STT Plugins Reference page linked from the table above. Writing a plugin instead of choosing one? See Writing an STT Plugin.