Instructions to use odinjason/koochik-fa-int8 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- NeMo
How to use odinjason/koochik-fa-int8 with NeMo:
import nemo.collections.asr as nemo_asr asr_model = nemo_asr.models.ASRModel.from_pretrained("odinjason/koochik-fa-int8") transcriptions = asr_model.transcribe(["file.wav"]) - Notebooks
- Google Colab
- Kaggle
koochik-fa-int8
Persian speech recognition for sherpa-onnx: an int8 copy of Shenava Koochik (NeMo FastConformer CTC, Persian) by Reza2kn.
model.onnx: the original model after dynamic int8 quantization of the MatMul weights (onnxruntimequantize_dynamic,QuantType.QInt8). This is the ONLY change; nothing was retrained. About 174 MB instead of about 459 MB.tokens.txt: unchanged from the original.
Use (Python, sherpa-onnx >= 1.13.7):
sherpa_onnx.OfflineRecognizer.from_nemo_ctc(model="model.onnx", tokens="tokens.txt")
Made for the Zolfstream app, which downloads this pack only when the user asks for it. Quality: tested by the publisher on two short Persian clips (informal and formal speech) against the unquantized model: near-identical output (one letter differed). This is NOT a benchmark.
License and credit: the original model is by Reza2kn and was reported under Apache-2.0; the same license applies to this modified copy. See the original model page for the full license text and any attribution requirements. No warranty.
Model tree for odinjason/koochik-fa-int8
Base model
nvidia/stt_fa_fastconformer_hybrid_large