SylReg-Decoder

This repository contains the decoder model presented in the paper Speaker-Disentangled Chunk-Wise Regression for Syllabic Tokenization.

• Project Page: https://ryota-komatsu.github.io/speaker_disentangled_hubert
• Repository: https://github.com/ryota-komatsu/speaker_disentangled_hubert

Model Details

Model Description

• Model type: Flow-matching-based Diffusion Transformer (DiT) with BigVGAN-v2
• Language(s) (NLP): English
• License: CC BY-NC-SA 4.0
• Finetuned from model: SylReg-Decoder Base

Model Sources

• Repository: Code
• Demo: Project page

How to Get Started with the Model

Use the code below to get started with the model.

git clone https://github.com/ryota-komatsu/speaker_disentangled_hubert.git
cd speaker_disentangled_hubert

sudo apt install git-lfs  # for UTMOS

conda create -y -n py310 -c pytorch -c conda-forge python=3.10 pip=24.0 setuptools=81.0.0 faiss-gpu=1.13.2 uv sox cuda-toolkit
conda activate py310
export UV_PROJECT_ENVIRONMENT=$CONDA_PREFIX
uv pip install -r requirements/requirements.txt

sh scripts/setup.sh

import re

import torch
import torchaudio
from datasets import Audio, load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer

from src.flow_matching import FlowMatchingWithBigVGan
from src.s5hubert.models.sylreg import SylRegForSyllableDiscovery

# download pretrained models from hugging face hub
encoder = SylRegForSyllableDiscovery.from_pretrained("ryota-komatsu/SylReg-Distill", device_map="cuda")
decoder = FlowMatchingWithBigVGan.from_pretrained("ryota-komatsu/SylReg-Decoder", device_map="cuda")

# load a waveform
dataset = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
dataset = dataset.cast_column("audio", Audio(sampling_rate=16000))
dataset = dataset.with_format("torch")
input_values = dataset[3]["audio"]["array"].unsqueeze(0)  # (1, T)

# encode a waveform into syllabic units
outputs = encoder(input_values.to(encoder.device))
units = outputs[0]["units"]  # [3950, 67, ..., 503]

# unit-to-speech synthesis
generated_speech = decoder(units.unsqueeze(0)).waveform.cpu()

torchaudio.save("input.wav", input_values, 16000)
torchaudio.save("output.wav", generated_speech, 16000)

Training Details

Training Data

Training Hyperparameters

• Training regime: fp16 mixed precision

Hardware

2 x A6000

Citation

@article{Komatsu_SylReg_2026,
  author    = {Komatsu, Ryota and Kawakita, Kota and Okamoto, Takuma and Shinozaki, Takahiro},
  title     = {Speaker-Disentangled Chunk-Wise Regression for Syllabic Tokenization},
  year      = {2026},
  volume    = {7},
  journal   = {IEEE Open Journal of Signal Processing},
  pages     = {800--808},
}