Post
966
Hi everyone!
I've been working on a pronunciation assessment engine optimized for edge deployment and real-time feedback. Wanted to share it with the community and get feedback.
**What it does**: Scores English pronunciation at 4 levels of granularity — phoneme, word, sentence, and overall (0-100 each). Returns IPA and ARPAbet notation for every phoneme.
**Key specs**:
- 17MB total model size (NeMo Citrinet-256, INT4 quantized)
- 257ms median inference on CPU
- Exceeds human inter-annotator agreement at phone-level (+4.5%) and sentence-level (+5.2%)
- Benchmarked on speechocean762 (2,500 test utterances)
- Tested across 7 L1 backgrounds (Chinese, Japanese, Korean, Arabic, Spanish, Vietnamese, Russian)
**Architecture**: CTC forced alignment + Viterbi decoding + GOP (Goodness of Pronunciation) scoring + MLP/XGBoost ensemble heads. No wav2vec2 dependency — the entire pipeline runs in 17MB.
**Try it**: fabiosuizu/pronunciation-assessment
The demo lets you record audio or upload a file, enter the expected text, and get instant scoring down to individual phonemes.
**API access**: Available via REST API, MCP servers (for AI agents), and Azure Marketplace. Details in the Space description.
Would love feedback on:
1. Use cases you'd find this useful for
2. Languages you'd want supported next
3. Whether the scoring feels calibrated for your experience level
Thanks!
I've been working on a pronunciation assessment engine optimized for edge deployment and real-time feedback. Wanted to share it with the community and get feedback.
**What it does**: Scores English pronunciation at 4 levels of granularity — phoneme, word, sentence, and overall (0-100 each). Returns IPA and ARPAbet notation for every phoneme.
**Key specs**:
- 17MB total model size (NeMo Citrinet-256, INT4 quantized)
- 257ms median inference on CPU
- Exceeds human inter-annotator agreement at phone-level (+4.5%) and sentence-level (+5.2%)
- Benchmarked on speechocean762 (2,500 test utterances)
- Tested across 7 L1 backgrounds (Chinese, Japanese, Korean, Arabic, Spanish, Vietnamese, Russian)
**Architecture**: CTC forced alignment + Viterbi decoding + GOP (Goodness of Pronunciation) scoring + MLP/XGBoost ensemble heads. No wav2vec2 dependency — the entire pipeline runs in 17MB.
**Try it**: fabiosuizu/pronunciation-assessment
The demo lets you record audio or upload a file, enter the expected text, and get instant scoring down to individual phonemes.
**API access**: Available via REST API, MCP servers (for AI agents), and Azure Marketplace. Details in the Space description.
Would love feedback on:
1. Use cases you'd find this useful for
2. Languages you'd want supported next
3. Whether the scoring feels calibrated for your experience level
Thanks!