SylReg-LM 7B

Model Details

Model Description

• Model type: Qwen2ForCausalLM
• Language(s) (NLP): English
• License: CC BY-NC-SA 4.0
• Finetuned from model: Qwen/Qwen2.5-7B

Model Sources

• Repository: Code
• Paper: arXiv:2607.04064
• 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
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 transformers import AutoModelForCausalLM, AutoTokenizer

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

wav_path = "/path/to/wav"

# download pretrained models from hugging face hub
encoder = SylRegForSyllableDiscovery.from_pretrained("ryota-komatsu/SylReg-Distill", device_map="cuda", dtype="auto")
decoder = FlowMatchingWithBigVGan.from_pretrained("ryota-komatsu/SylReg-Decoder", device_map="cuda", dtype="auto")
speechlm = AutoModelForCausalLM.from_pretrained("ryota-komatsu/SylReg-LM-7B", device_map="cuda", dtype="auto")
tokenizer = AutoTokenizer.from_pretrained("ryota-komatsu/SylReg-LM-7B")

# load a waveform
waveform, sr = torchaudio.load(wav_path)
waveform = torchaudio.functional.resample(waveform, sr, 16000)

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

# speech language modeling
text = "".join(f"<{unit}>" for unit in units)
input_ids = tokenizer(text, padding=True, return_tensors="pt").input_ids.to(speechlm.device)
generated_ids = speechlm.generate(input_ids=input_ids, do_sample=True, temperature=0.8)[0]
units = tokenizer.decode(generated_ids)
units = torch.tensor([int(unit) for unit in re.findall(r"<(\d+)>", units)], device=decoder.device)

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

Training Details

Training Data

Training Hyperparameters

• Training regime: bf16 mixed precision
• Training steps: 15k
• Batch size: 2,097,152 (=2²¹) tokens
• Optimizer: AdamW(lr=0.0003, betas=(0.9, 0.95), weight_decay=0.01)
• Scheduler: warmup_stable_decay(num_warmup_steps=100, num_decay_steps=5000, min_lr_ratio=0.1)

Hardware

32 NVIDIA H100 GPUs

Citation

BibTeX:

@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},
}