Practical overview
When it fits
Use Soxan when you want research-oriented examples and scripts for adapting pretrained speech models to classification tasks, rather than a hosted speech-to-text service.
Who it is for
- Researchers exploring speech emotion recognition or audio classification
- Developers adapting pretrained speech models to labeled datasets
Documented uses
- Train a speech emotion or audio classifier using dataset files and a pretrained model.
- Explore notebook examples for Wav2Vec 2.0 and HuBERT.
- Apply a saved speech-classification model to an audio file and inspect its label scores.
Requirements & limits
The public GitHub repository is licensed under Apache-2.0.
- The documented training workflow expects train, validation and test CSV files, but the README does not describe their schema.
- The repository includes notebook examples for Greek speech emotion recognition and eating-sound classification.
- The README lists Persian emotion and gender recognition models and a Greek emotion-recognition model.
- The README labels its Persian speech-to-text demo as “soon.”
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Train a speech classifier from CSV datasets
Configure the documented command-line training script to use a pretrained model and dataset files.
Before you start
- A pretrained model identifier or path
- Training, validation and test CSV files
- Output and cache directories
- Python 3
Steps
Set the model identifier, selecting Wav2Vec 2.0 or HuBERT through the documented model options.
Provide paths for the training, validation and test CSV files, plus output and cache directories.
Set the documented training options, including pooling mode, batch sizes, learning rate and number of epochs.
Use the script’s training and evaluation flags as shown in the README.
Expected outcome
A configured command for training and evaluating a speech classification model on the supplied dataset.
Guide limitations
- The README does not specify the CSV schema.
- The example command contains placeholder paths and is not a guarantee of a successful run.
Apply a saved model to an audio file
Use the documented Python prediction example to produce emotion scores for an audio file.
Before you start
- A saved Wav2Vec 2.0 or HuBERT speech-classification model
- An audio file
- PyTorch, torchaudio and Transformers
Steps
Load the model configuration and feature extractor from the model path.
Load the speech-classification model and select CUDA when available or CPU otherwise.
Load the audio file and resample it to the feature extractor’s sampling rate.
Run the model on the extracted inputs and map output scores to the configured labels.
Expected outcome
A list of emotion labels with corresponding scores for the input audio.
Guide limitations
- The prediction example is for speech classification and emotion labels, not general-purpose speech transcription.
- The README uses a placeholder model path and audio path.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 15 October 2026
- Original website title
- m3hrdadfi/soxan · Official homepage title · retained in its source language