PL-Flow pretrained bundle

Model weights for the PL-Flow two-stage speech synthesis implementation. Licenses differ by component; see THIRD_PARTY.md.

Components

Component	Function	Origin	Weight license
s1	Phoneme-aligned prosody flow, QK normalization	Trained from random initialization, EMA	GPL-3.0-only
s2	Duration-conditioned acoustic flow, QK normalization	Trained from random initialization, EMA	GPL-3.0-only
aligner	HuBERT/text forced alignment	Trained from random initialization	GPL-3.0-only
vocoder	32 kHz mono, 32-channel 50 Hz SQ-GAN codec	Trained from random initialization, EMA	GPL-3.0-only
plbert	Frozen phoneme ALBERT	Extracted from Kokoro	Apache-2.0
hubert	Frozen semantic features	TencentGameMate Chinese HuBERT	MIT
speaker	Frozen WavLM/ECAPA speaker features	External checkpoint linked by Seed-TTS-Eval	Download and upstream terms

Six components contain config.json and model.safetensors. The speaker directory contains configuration and download instructions only. bundle.json pins SHA256 hashes, the external speaker download, the prosody-code interface and inference settings.

Usage

Install the PL-Flow source package in your own environment, then:

hf download rcell233/pl-flow --local-dir artifacts/pretrained

Download wavlm_large_finetune.pth from the model link in Seed-TTS-Eval and save it as artifacts/pretrained/speaker/wavlm_large_finetune.pth. PL-Flow verifies the checkpoint's SHA256 and extracts speaker features automatically during synthesis.

pl-flow verify --models artifacts/pretrained
pl-flow synthesize --models artifacts/pretrained \
  --reference-audio data/reference.wav --reference-text '这是参考音频的文本。' \
  --text '你好，欢迎使用语音合成。' --language zh --output outputs/example.wav

Default sampling uses S1 8 Euler steps / CFG 2.1, S2 10 Euler steps / CFG 2.8, temperature 1.0 and CUDA BF16 autocast for the flow models. Guidance is (1 + scale) * conditional - scale * unconditional. Audio encoders and the codec run in FP32. Context limits: 512 S1 tokens, 1024 total S2 frames, 250 prompt frames.

S1 and S2 use EMA weights. The bundle supports inference and initialization for training. Resuming training requires a full training checkpoint. The ASR aligner and SQ codec can also be trained using the PL-Flow package.