storyteller-platform/quartznet15x5-it

QuartzNet 15x5 Italian

Storyteller installs this model through ghost-story. This repository mirrors version 2.0.0 of quartznet15x5-it in the Storyteller model registry, with a folder per format.

Trained by NVIDIA, nvidia/stt_it_quartznet15x5, licensed CC-BY-4.0. Converted from the ONNX export OpenVoiceOS/stt_it_quartznet15x5_onnx at revision 178fb4be74ac.

• model.safetensors holds the encoder weights in the layout the MLX loader expects, plus the mel filterbank and window of the NeMo front end.
• model.onnx is a new graph that export-onnx.py builds from those weights. The mel front end is part of it, so it takes the pre-emphasised, reflect padded signal instead of features: the short-time Fourier transform runs as a convolution, followed by the mel projection, the log and the per-feature normalisation. That way it runs on every onnxruntime execution provider without a separate feature extractor.
• model_fp16.onnx is the same graph with the encoder in half precision. The front end stays in single precision.
• config.json describes the architecture, the front end and the block layout, written by convert.py.
• vocab.txt is unchanged from the export.

The onnx format holds model.onnx and model_fp16.onnx, the mlx format model.safetensors, each with the configuration and vocabulary.

convert.py and export-onnx.py are in libraries/mlx-audio/scripts.