sherpa-onnx-kokoro-ax650

Kokoro TTS model converted to AX650 AXMODEL for on-device text-to-speech inference via sherpa-onnx.

Model Details

Original Model: hexgrad/Kokoro-82M
Architecture: KModelForONNX with conv-based ISTFT
Task: Text-to-Speech (TTS)
Sample rate: 24000 Hz
Voices: af_heart (1 American English voice)
Target Chip: AX650 (NPU3)

Files

File	Size	Description
`models/kokoro_part1_96.axmodel`	86.8 MB	Kokoro TTS encoder part 1
`models/kokoro_part2_96.axmodel`	21.4 MB	Kokoro TTS encoder part 2
`models/kokoro_part3_96.axmodel`	67.7 MB	Kokoro TTS decoder
`models/model4_har_sim.onnx`	14 KB	Reference ONNX model
`checkpoints/config.json`	2.3 KB	Model configuration
`checkpoints/voices_npy/af_heart.npy`	510 KB	Voice embedding (af_heart)
`cpp/dict/vocab.txt`	801 B	Vocabulary file

Usage with sherpa-onnx

./sherpa-onnx-offline-tts \
  --kokoro-model=./models/kokoro_part1_96.axmodel \
  --kokoro-model2=./models/kokoro_part2_96.axmodel \
  --kokoro-model3=./models/kokoro_part3_96.axmodel \
  --kokoro-voices=./checkpoints/voices_npy/af_heart.npy \
  --kokoro-tokens=./cpp/dict/vocab.txt \
  --kokoro-data-dir=./checkpoints \
  --output-filename=./output.wav \
  "Hello world!"