See My Voice Mandarin Phone CTC
This is the trained Mandarin phone-token CTC model used by the See My Voice pronunciation practice app.
The model maps Mandarin speech audio to the app's 67 pronunciation labels:
audio -> initial/final/tone phone tokens
Example target labels:
你 ni2 -> I_n F_i T2
好 hao3 -> I_h F_ao T3
是 shi4 -> I_sh F_i_zh T4
字 zi4 -> I_z F_i_z T4
儿 er2 -> F_er T2
女 nv3 -> I_n F_v T3
Training Summary
- Base model:
jonatasgrosman/wav2vec2-large-xlsr-53-chinese-zh-cn - Output labels: 67 Mandarin phone tokens
- Training device: CUDA
- Training examples: 300
- Validation examples: 50
- Max steps: 1000
- Final training loss: 1.5815
- Best validation loss: 1.9028
- Best validation step: 1000
The training run kept the large acoustic encoder frozen and trained the CTC output head for the See My Voice phone-token target space.
Intended Use
This model is intended for the See My Voice pronunciation feedback service. The service compares expected Mandarin phone tokens for a target sentence against the model's predicted phone tokens from learner audio.
Limitations
This is an early project model, not a production-grade Mandarin pronunciation assessment model. It was trained on a small dataset and should be treated as a prototype signal for initial/final/tone feedback rather than a definitive linguistic or educational assessment.
See training_report.json, loss_history.csv, and test_loss_report.json for
the generated run details.