Embed2ASR v0.1.0

Embed2ASR v0.1.0 release notes

Summary

First research release of Embed2ASR: speech recognition built on Google's EmbeddingGemma 2 audio encoder, kept frozen, with a 9.66M-parameter GRU-CTC reader trained on top. Parameters loaded for inference: ~314.5M. Languages: Taiwanese Mandarin, English, and Mandarin-English code-switching.

What's included

• Inference package (greedy CTC; optional Chinese character 5-gram LM).
• Training code for the reader.
• Benchmark script and results: benchmarks.md, data in benchmarks/
• Documentation: architecture, method, limitations, reproducibility, roadmap.

Benchmark

Test-set results, speed and memory: benchmarks.md. CER, WER and MER are reported separately; greedy and LM separately.

Weights

Hugging Face: https://huggingface.co/tc3oliver/Embed2ASR-v1

Embed2ASR-v1 artifacts (reader weights, tokenizer, unit list, optional LM) are released under Apache-2.0. Checkpoint: Embed2ASR-v1, epoch 74, sha256 90d1861dde754574ba0d8cc4a116fb7a7a7a5bf94d48a72c474becc2798aee29. The EmbeddingGemma 2 encoder is not included; download it from Google.

Known limitations

See limitations.md. Highlights: frozen-encoder ceiling, homophone errors from CTC, English behind Whisper large-v3, narrow code-switching evidence, offline only.

Credits

EmbeddingGemma 2 was created by Google.