Model Card: SpeechAligner Reference Model
May 25, 2026 ยท View on GitHub
This is a lightweight model card for the reference model in
burn-speech-training. It is meant to make the current limits clear, not to
present the model as production-ready.
Model
- Name:
SpeechAligner - Framework: Burn 0.21
- Default shape: CNN + squeeze-and-excitation blocks + self-attention + CTC head
- Default size: about 1.7M parameters
- Quickstart size: about 122K parameters
Intended Use
- Learning how to wire a speech training loop in Rust with Burn.
- Inspecting a small CTC-based alignment path.
- Reusing pieces such as feature extraction, batching, model structure, or experiment logging.
- Running controlled experiments on pronunciation scoring ideas.
Not Intended For
- Production ASR.
- Certified language assessment.
- Medical, hiring, immigration, or high-stakes scoring.
- Claims about learner ability without further validation.
- Real-time inference guarantees.
Training And Evaluation Data
- Synthetic quickstart data: smoke test only.
- LibriSpeech: used by the real-data training path.
- SpeechOcean762: used for pronunciation scoring experiments.
- CMUdict: used for transcript to phoneme lookup.
See docs/datasets.md for provenance and inclusion notes.
Reported Result
The best reported pronunciation-scoring result in the experiment log is about
0.292 Spearman correlation against SpeechOcean762 word-level human scores.
That number is a limitation marker, not a victory claim. The evidence points to representation and data limits more than another small loss-function tweak.
Known Failure Modes
- MFCC/log-mel features may be too weak for robust pronunciation ranking.
- CTC alignment loss can interfere with scoring fine-tuning in this setup.
- SpeechOcean762 word scores are heavily imbalanced toward perfect scores.
- Out-of-vocabulary words depend on fallback phoneme handling.
- GPU features are hardware-dependent and not fully covered by CI.
Reproducibility Notes
For a quick smoke test:
cargo run --example train_small --features ndarray --release
For experiment claims, prefer linking to docs/experiment-log.md and include
the dataset split, command, commit, hardware, and metric definition.