BitVoice ASR models
On-device speech-recognition models for the BitVoice dictation app, run via
CRISP ASR (dart:ffi + ggml).
BitVoice has no in-app networking: download a model here in your browser, then
import the file into the app (Model manager → Import). Each model is a
single .gguf/.bin; Moonshine is a .zip (model + tokenizer) the app
unpacks on import.
See SOURCES.md for the upstream repo of each model (where to pull the next release from).
Recommended (small, fast, tested)
| Model | Lang | Size | Notes | Download |
|---|---|---|---|---|
| SenseVoice small | EN/ZH/JA/KO/yue | 129 MB | Best accuracy, multilingual | sensevoice-small-q4.gguf |
| Whisper base | EN | 57 MB | Reliable default (1.5% WER) | whisper-base-en-q5.bin |
| Parakeet 110M | EN | 75 MB | Fastest English | parakeet-110m-en-q4.gguf |
| Moonshine tiny | EN | 19 MB | Smallest & fastest | moonshine-tiny-en-q4.zip |
More architectures (advanced)
| Model | Lang | Size | Notes | Download |
|---|---|---|---|---|
| Paraformer | ZH | 123 MB | Fast Chinese ASR | paraformer-zh-q4.gguf |
| wav2vec2 XLSR | EN | 212 MB | CTC (no punctuation) | wav2vec2-xlsr-en-q4.gguf |
| GigaAM v3 | RU | 237 MB | Russian, punctuation+casing | gigaam-v3-ru-q8.gguf |
| Nemotron 3.5 (streaming) | EN | 390 MB | NVIDIA 0.6B | nemotron-streaming-en-q4.gguf |
| Parakeet 0.6B v3 | EN | 399 MB | Higher accuracy | parakeet-0.6b-en-q4.gguf |
| Canary 1B v2 | EN/DE/ES/FR | 583 MB | Multilingual, high-RAM | canary-1b-q4.gguf |
| Qwen3-ASR 0.6B | multi | 602 MB | LLM-decoder, high-RAM | qwen3-asr-q4.gguf |
| ReazonSpeech v2 | JA | 636 MB | Japanese, high-RAM | reazonspeech-ja-q8.gguf |
Larger models (Canary/Qwen3/Reazon) need a high-RAM phone (≈6 GB+).
Provenance & attribution
Re-hosted quantized conversions. GGUF conversions by cstr (except Whisper, from ggerganov/whisper.cpp). Original credits and licenses:
| Model | Origin | License |
|---|---|---|
| Whisper base.en | OpenAI / whisper.cpp | MIT |
| Moonshine tiny | UsefulSensors/moonshine | MIT |
| SenseVoice small | FunAudioLLM/SenseVoiceSmall | funasr (commercial OK w/ attribution) |
| Paraformer zh | FunASR | funasr (commercial OK w/ attribution) |
| Parakeet 110M / 0.6B v3 | nvidia/parakeet-* | CC-BY-4.0 |
| wav2vec2 XLSR-53 en | facebook/wav2vec2-large-xlsr-53 | Apache-2.0 |
| GigaAM v3 | ai-sage/GigaAM-v3 | MIT |
| Nemotron 3.5 ASR streaming | NVIDIA | see NVIDIA model card |
| Qwen3-ASR 0.6B | Qwen/Qwen3-ASR | Apache-2.0 |
| Canary 1B v2 | nvidia/canary-1b-v2 | CC-BY-4.0 |
| ReazonSpeech NeMo v2 | reazon-research | Apache-2.0 |
CC-BY-4.0 models are redistributed with attribution as required. Consult each original model card for authoritative terms.
- Downloads last month
- -
We're not able to determine the quantization variants.