| Sado is the name AlgorithVoice gives to STT Uz FastConformer Hybrid | |
| Transducer-CTC Large P&C by NVIDIA Corporation, licensed under CC-BY-4.0 | |
| (https://creativecommons.org/licenses/by/4.0/). NVIDIA does not endorse | |
| AlgorithVoice or this repository. | |
| This repository is a copy of nvidia/stt_uz_fastconformer_hybrid_large_pc with | |
| these files added, for use with sherpa-onnx in AlgorithVoice: | |
| encoder.int8.onnx decoder.onnx joiner.onnx tokens.txt | |
| Changes, as CC-BY asks to be stated: the files are the transducer (RNNT) | |
| branch of the same model, exported to ONNX by Algorithco with NeMo, with the | |
| metadata sherpa-onnx reads. The encoder's MatMul weights are quantized to 8 | |
| bits; the convolutions, the decoder and the joiner stay in 32-bit float, | |
| because a quantized joiner drops words. Nothing was trained. | |
| Measured on 100 FLEURS uz_uz test recordings: 14.1% word error rate with | |
| these files in AlgorithVoice, against 14.4% for the CTC branch. | |