Instructions to use ahmedmohamedra2002/streamvoice-speech-models with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- NeMo
How to use ahmedmohamedra2002/streamvoice-speech-models with NeMo:
import nemo.collections.asr as nemo_asr asr_model = nemo_asr.models.ASRModel.from_pretrained("ahmedmohamedra2002/streamvoice-speech-models") transcriptions = asr_model.transcribe(["file.wav"]) - Notebooks
- Google Colab
- Kaggle
StreamVoice speech models
Small speech recognition models of one language each, used by the StreamVoice Android app ("small precise" models). They are NVIDIA's FastConformer models (115M parameters, with punctuation and capitals), converted to the app's ggml format and quantised to 5 bits (q5_0). Each file serves one language only.
| File | Language | Source model (NVIDIA, CC-BY-4.0) |
|---|---|---|
| streamvoice-asr-ar-q5_0.bin | Arabic | nvidia/stt_ar_fastconformer_hybrid_large_pc_v1.0 |
| streamvoice-asr-fr-q5_0.bin | French | nvidia/stt_fr_fastconformer_hybrid_large_pc |
| streamvoice-asr-nl-q5_0.bin | Dutch | nvidia/stt_nl_fastconformer_hybrid_large_pc |
| streamvoice-asr-pl-q5_0.bin | Polish | nvidia/stt_pl_fastconformer_hybrid_large_pc |
| streamvoice-asr-hr-q5_0.bin | Croatian | nvidia/stt_hr_fastconformer_hybrid_large_pc |
| streamvoice-asr-be-q5_0.bin | Belarusian | nvidia/stt_be_fastconformer_hybrid_large_pc |
| streamvoice-asr-ka-q5_0.bin | Georgian | nvidia/stt_ka_fastconformer_hybrid_large_pc |
| streamvoice-asr-es-q5_0.bin | Spanish | nvidia/stt_es_fastconformer_hybrid_large_pc |
| streamvoice-asr-de-q5_0.bin | German | nvidia/stt_de_fastconformer_hybrid_large_pc |
| streamvoice-asr-ru-q5_0.bin | Russian | nvidia/stt_ru_fastconformer_hybrid_large_pc |
| streamvoice-asr-it-q5_0.bin | Italian | nvidia/stt_it_fastconformer_hybrid_large_pc |
Changes made to the original models
- Converted from NeMo to the ggml format the app's speech engine reads; the CTC head of the hybrid models is not used (the transducer decoder is). The transducer's joint output gets one extra all-zero row (a duration of 0), so the app's TDT decoder can run it; the encoder's input scaling is folded into its first layer.
- Weights quantised to 5 bits (q5_0); 4 bits was not used (it broke these models).
- Nothing was retrained.
Licence
Each model is ยฉ NVIDIA, licensed under Creative Commons Attribution 4.0 (CC-BY-4.0). These files are adaptations of the models listed above. NVIDIA's Portuguese model is licensed for non-commercial use only and is therefore not included; the Ukrainian and Persian ones were measured less accurate than other small models and are not included either.
- Downloads last month
- -