Piper VITS en_US Lessac Medium (Ambarella N1-655 CVflow v3 NVP)

Piper VITS (en_US-lessac-medium) is a neural text-to-speech (TTS) model based on VITS (Variational Inference with adversarial learning for end-to-end Text-to-Speech), partitioned into a static-shape phoneme encoder (vits_enc.bin) and a 50-frame overlap-tiled waveform decoder (vits_dec.bin) for the Ambarella N1-655 CVflow v3 NVP.

Non-Commercial Dataset Notice (blizzard-2013-non-commercial): While the Piper VITS model architecture and code are licensed under the MIT License, the en_US-lessac-medium voice checkpoint was trained on the Blizzard Challenge 2013 (Lessac Technologies, Inc.) dataset. Under the Blizzard Challenge 2013 dataset license terms, these voice weights are restricted to non-commercial research, educational, and demonstration use only and may not be used for commercial purposes without separate authorization from Lessac Technologies, Inc.

Original paper: Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech

Model Configuration:

• Reference implementation: https://github.com/rhasspy/piper
• Original Weight: https://huggingface.co/rhasspy/piper-voices/tree/main/en/en_US/lessac/medium
• Dataset: Blizzard Challenge 2013 — Lessac Technologies, Inc. (Non-Commercial Research License)
• License: other (blizzard-2013-non-commercial, see LICENSE — MIT code + Blizzard Challenge 2013 Non-Commercial Dataset)
• Support Cooper version:
  • CNNGen: [3.9.0 / n1-655x] (cavalry_gen 3.0.4)

Upstream Attribution & Modification Notice

• Upstream Code & Architecture Copyright: Copyright (c) 2023 Michael Hansen (rhasspy / OHF-Voice piper), licensed under the MIT License.
• Upstream Voice Dataset License (blizzard-2013-non-commercial): The en_US-lessac-medium voice weights are derived from the Blizzard Challenge 2013 corpus provided by Lessac Technologies, Inc. and distributed for non-commercial research, educational, and evaluation use only.
• Modification Notice: Modified by Ambarella International LLC (2026). Partitioned at the stochastic duration predictor / length regulator into Stage 0 (vits_enc.bin) and Stage 1 50-frame overlap-tiled HiFi-GAN decoder (vits_dec.bin), 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: vits_enc.bin, vits_dec.bin, phoneme_id_map.json
• Stage 0 (vits_enc.bin): input [1, 128] int16, x_mask [1, 1, 128] FP16 -> m_p [1, 192, 128], logs_p [1, 192, 128], logw [1, 1, 128] IEEE FP16
• CPU Bridge: d_i = ceil(exp(logw_i) * x_mask_i * length_scale), expand m_p/logs_p to [1, 192, 500] (y_mask [1, 1, 500])
• Stage 1 (vits_dec.bin): 50-frame overlap-tiled (core=20, ctx=15) z_p [1, 192, 50], y_mask [1, 1, 50] FP16 -> audio [1, 1, 1, 12800] IEEE FP16 @ 22.05 kHz (hop=256)

Model	Device	Compression	Model Link
Piper	N1-655	Amba_optimized	Model_Link