hama on-device NLP
← Back to blog

2026. 7. 19.

hama v1.7.0: v3 phoneme ASR

Version 1.7.0 refreshes hama's shipped phoneme ASR runtime. The public API stays the same, but the packaged model is now the v3 epoch-3 EMA Conformer from hama-training, and the default decode temperature matches the training-aligned setting used in evaluation.

  • Replaced the shipped v2 phoneme ASR encoder with the v3 epoch-3 EMA Conformer. The waveform ABI, 191-token vocabulary, 320-sample output-frame timing, and public APIs are unchanged.
  • The new encoder has 2,005,535 parameters and four 144-channel Conformer blocks. The Zig engine now carries the full v3 forward path directly for native and WebAssembly inference, including macaron feed-forwards, four-head self-attention, and GLU/depthwise-convolution modules.
  • Updated the default ASR decode temperature from 1.0 to 0.95 in Python, Node, and the browser. The blank bias remains -0.1.
  • Epoch-3 EMA PER is 22.08% on the 36,359-row robust test and 22.60% on the legacy 5,559-row test. The LibriSpeech slice is 14.76% overall (clean 10.94%, other 18.71%).
  • Regenerated ASR assets, stage fixtures, native libraries, hama.wasm, and public goldens. Validation passes 31 Zig tests, 86 Python tests, and 85 TypeScript/WASM tests.