cchiou-amba/MatchaTTS

Matcha-TTS LJSpeech (Ambarella N1-655 CVflow v3 NVP)

Matcha-TTS is a fast conditional flow-matching acoustic text-to-speech architecture trained on LJSpeech-1.1 (Public Domain). It is partitioned at the duration/length regulator into a static-shape phoneme encoder (matcha_enc.bin) and a flow-matching ODE decoder (matcha_dec.bin) that outputs 80-bin log-Mel spectrograms ([1, 80, 500]) for synthesis with HiFiGAN on the Ambarella N1-655 CVflow v3 NVP.

Original paper: Matcha-TTS: A fast TTS architecture with conditional flow matching

Model Configuration:

• Reference implementation: https://github.com/shivammehta25/Matcha-TTS
• Original Weight: https://huggingface.co/csukuangfj/matcha-icefall-en_US-ljspeech
• Dataset: LJSpeech-1.1 (Public Domain)
• License: Apache-2.0 / MIT (LICENSE, trained on Public Domain LJSpeech-1.1)
• Support Cooper version:
  • CNNGen: [3.9.0 / n1-655x] (cavalry_gen 3.0.4)

Upstream Attribution & Modification Notice

• Upstream Copyright: Matcha-TTS architecture Copyright (c) 2023 Shivam Mehta (shivammehta25/Matcha-TTS, MIT License); k2-fsa/icefall LJSpeech recipe and ONNX export Copyright (c) k2-fsa/icefall & sherpa-onnx contributors (csukuangfj/matcha-icefall-en_US-ljspeech, Apache License 2.0); trained on the LJSpeech-1.1 dataset (Public Domain).
• Modification Notice (Apache-2.0 Section 4(b)): Modified by Ambarella International LLC (2026). Partitioned at the duration predictor / length regulator into Stage 0 (matcha_enc.bin, [1, 128] -> mu_x [1, 80, 128], logw [1, 1, 128]) and Stage 1 (matcha_dec.bin, mu_y [1, 80, 500] -> mel [1, 80, 500]), and compiled via Ambarella Cooper SDK CNNGen 3.9.0 (cavalry_gen 3.0.4) for CVflow v3 N1-655 (vdg_arch = 0xB).

NVP Port & Runtime Specification

• Target SoC: Ambarella N1-655 (vdg_arch = 0xB, CVflow v3 cavalry_gen 3.0.4)
• Archive members: matcha_enc.bin, matcha_dec.bin, tokens.txt
• Stage 0 (matcha_enc.bin): x [1, 128] int16, x_mask [1, 1, 128] FP16 -> mu_x [1, 80, 128], logw [1, 1, 128] IEEE FP16
• CPU Bridge: d_i = ceil(exp(logw_i) * x_mask_i * length_scale), expand mu_x to mu_y [1, 80, 500] and y_mask [1, 1, 500]
• Stage 1 (matcha_dec.bin): mu_y [1, 80, 500], y_mask [1, 1, 500], z [1, 80, 500] FP16 -> mel [1, 80, 500] IEEE FP16 (chains into HiFi-GAN)

Model	Device	Compression	Model Link
MatchaTTS	N1-655	Amba_optimized	Model_Link