P-Flow: A Fast and Data-Efficient Zero-Shot TTS through Speech Prompting

Authors : Sungwon Kim, Kevin J Shih, Rohan Badlani, Joao Felipe Santos, Evelina Bhakturina, Mikyas Desta, Rafael Valle, Sungroh Yoon, Bryan Catanzaro

Affiliations: NVIDIA

Status : Added newer samples with better prosody and pronunciation. Check out `samples` folder. LJSpeech pretrained ckpt - GDrive Link

Unofficial implementation of the paper P-Flow: A Fast and Data-Efficient Zero-Shot TTS through Speech Prompting by NVIDIA.

While recent large-scale neural codec language models have shown significant improvement in zero-shot TTS by training on thousands of hours of data, they suffer from drawbacks such as a lack of robustness, slow sampling speed similar to previous autoregressive TTS methods, and reliance on pre-trained neural codec representations. Our work proposes P-Flow, a fast and data-efficient zero-shot TTS model that uses speech prompts for speaker adaptation. P-Flow comprises a speechprompted text encoder for speaker adaptation and a flow matching generative decoder for high-quality and fast speech synthesis. Our speech-prompted text encoder uses speech prompts and text input to generate speaker-conditional text representation. The flow matching generative decoder uses the speaker-conditional output to synthesize high-quality personalized speech significantly faster than in real-time. Unlike the neural codec language models, we specifically train P-Flow on LibriTTS dataset using a continuous mel-representation. Through our training method using continuous speech prompts, P-Flow matches the speaker similarity performance of the large-scale zero-shot TTS models with two orders of magnitude less training data and has more than 20× faster sampling speed. Our results show that P-Flow has better pronunciation and is preferred in human likeness and speaker similarity to its recent state-of-the-art counterparts, thus defining P-Flow as an attractive and desirable alternative.

Credits

Of course the kind author of the paper for taking some time to explain me some details of the paper that I didn't understand at first.
We will build this repo based on the VITS2 repo, MATCHA-TTS repo and VoiceFlow-TTS repo

Dry run

cd pflowtts_pytorch/notebooks

import sys
sys.path.append('..')

from pflow.models.pflow_tts import pflowTTS
import torch
from dataclasses import dataclass

@dataclass
class DurationPredictorParams:
    filter_channels_dp: int
    kernel_size: int
    p_dropout: float

@dataclass
class EncoderParams:
    n_feats: int
    n_channels: int
    filter_channels: int
    filter_channels_dp: int
    n_heads: int
    n_layers: int
    kernel_size: int
    p_dropout: float
    spk_emb_dim: int
    n_spks: int
    prenet: bool

@dataclass
class CFMParams:
    name: str
    solver: str
    sigma_min: float

# Example usage
duration_predictor_params = DurationPredictorParams(
    filter_channels_dp=256,
    kernel_size=3,
    p_dropout=0.1
)

encoder_params = EncoderParams(
    n_feats=80,
    n_channels=192,
    filter_channels=768,
    filter_channels_dp=256,
    n_heads=2,
    n_layers=6,
    kernel_size=3,
    p_dropout=0.1,
    spk_emb_dim=64,
    n_spks=1,
    prenet=True
)

cfm_params = CFMParams(
    name='CFM',
    solver='euler',
    sigma_min=1e-4
)

@dataclass
class EncoderOverallParams:
    encoder_type: str
    encoder_params: EncoderParams
    duration_predictor_params: DurationPredictorParams

encoder_overall_params = EncoderOverallParams(
    encoder_type='RoPE Encoder',
    encoder_params=encoder_params,
    duration_predictor_params=duration_predictor_params
)

@dataclass
class DecoderParams:
    channels: tuple
    dropout: float
    attention_head_dim: int
    n_blocks: int
    num_mid_blocks: int
    num_heads: int
    act_fn: str

decoder_params = DecoderParams(
    channels=(256, 256),
    dropout=0.05,
    attention_head_dim=64,
    n_blocks=1,
    num_mid_blocks=2,
    num_heads=2,
    act_fn='snakebeta',
)
    
model = pflowTTS(
    n_vocab=100,
    n_feats=80,
    encoder=encoder_overall_params,
    decoder=decoder_params.__dict__,
    cfm=cfm_params,
    data_statistics=None,
)

x = torch.randint(0, 100, (4, 20))
x_lengths = torch.randint(10, 20, (4,))
y = torch.randn(4, 80, 500)
y_lengths = torch.randint(300, 500, (4,))

dur_loss, prior_loss, diff_loss, attn = model(x, x_lengths, y, y_lengths)
# backpropagate the loss 

# now synthesises
x = torch.randint(0, 100, (1, 20))
x_lengths = torch.randint(10, 20, (1,))
y_slice = torch.randn(1, 80, 264)

model.synthesise(x, x_lengths, y_slice, n_timesteps=10)

Quick run in Google Colab

Instructions to run

Create an environment (suggested but optional)

conda create -n pflowtts python=3.10 -y
conda activate pflowtts

Stay in the root directory (of course clone the repo first!)

cd pflowtts_pytorch
pip install -r requirements.txt

Build Monotonic Alignment Search

# Cython-version Monotonoic Alignment Search
python setup.py build_ext --inplace

Let's assume we are training with LJ Speech

Download the dataset from here, extract it to data/LJSpeech-1.1, and prepare the file lists to point to the extracted data like for item 5 in the setup of the NVIDIA Tacotron 2 repo.

3a. Go to configs/data/ljspeech.yaml and change

train_filelist_path: data/filelists/ljs_audio_text_train_filelist.txt
valid_filelist_path: data/filelists/ljs_audio_text_val_filelist.txt

3b. Helper commands for the lazy

!mkdir -p /home/ubuntu/LJSpeech/LJSpeech-1.1/filelists
!wget -O /home/ubuntu/LJSpeech/LJSpeech-1.1/filelists/ljs_audio_text_test_filelist.txt https://raw.githubusercontent.com/NVIDIA/tacotron2/master/filelists/ljs_audio_text_test_filelist.txt
!wget -O /home/ubuntu/LJSpeech/LJSpeech-1.1/filelists/ljs_audio_text_train_filelist.txt https://raw.githubusercontent.com/NVIDIA/tacotron2/master/filelists/ljs_audio_text_train_filelist.txt
!wget -O /home/ubuntu/LJSpeech/LJSpeech-1.1/filelists/ljs_audio_text_val_filelist.txt https://raw.githubusercontent.com/NVIDIA/tacotron2/master/filelists/ljs_audio_text_val_filelist.txt

!sed -i -- 's,DUMMY,/home/ubuntu/LJSpeech/LJSpeech-1.1/wavs,g' /home/ubuntu/LJSpeech/LJSpeech-1.1/filelists/*.txt

!sed -i -- 's,train_filelist_path: data/filelists/ljs_audio_text_train_filelist.txt,train_filelist_path: /home/ubuntu/LJSpeech/LJSpeech-1.1/filelists/ljs_audio_text_train_filelist.txt,g' /home/ubuntu/LJSpeech/pflowtts_pytorch/configs/data/ljspeech.yaml
!sed -i -- 's,valid_filelist_path: data/filelists/ljs_audio_text_val_filelist.txt,valid_filelist_path: /home/ubuntu/LJSpeech/LJSpeech-1.1/filelists/ljs_audio_text_val_filelist.txt,g' /home/ubuntu/LJSpeech/pflowtts_pytorch/configs/data/ljspeech.yaml

Generate normalisation statistics with the yaml file of dataset configuration

cd pflowtts_pytorch/pflow/utils
python generate_data_statistics.py -i ljspeech.yaml
# Output:
#{'mel_mean': -5.53662231756592, 'mel_std': 2.1161014277038574}

Update these values in configs/data/ljspeech.yaml under data_statistics key.

data_statistics:  # Computed for ljspeech dataset
  mel_mean: -5.536622
  mel_std: 2.116101

to the paths of your train and validation filelists.

Run the training script

python pflow/train.py experiment=ljspeech

for multi-gpu training, run

python pflow/train.py experiment=ljspeech trainer.devices=[0,1]

Architecture details

Speech prompted text encoder with Prenet and RoPE Transformer
Duration predictor with MAS
Flow matching generative decoder with CFM (paper uses wavenet decoder; we use modified wavenet and optional U-NET decoder is included to experiment with)
Speech prompt input currently slices the input spectrogram and concatenates it with the text embedding. Can support external speech prompt input (during training as well)
pflow prompt masking loss for training
HiFiGan for vocoder
Guidance for sampling

smksyj/pflowtts_pytorch

P-Flow: A Fast and Data-Efficient Zero-Shot TTS through Speech Prompting

Authors : Sungwon Kim, Kevin J Shih, Rohan Badlani, Joao Felipe Santos, Evelina Bhakturina, Mikyas Desta, Rafael Valle, Sungroh Yoon, Bryan Catanzaro

Affiliations: NVIDIA

Status : Added newer samples with better prosody and pronunciation. Check out `samples` folder. LJSpeech pretrained ckpt - GDrive Link

Credits

Dry run

Quick run in Google Colab

Instructions to run

Architecture details

TODOs, features and update notes

smksyj/pflowtts_pytorch

P-Flow: A Fast and Data-Efficient Zero-Shot TTS through Speech Prompting

Authors : Sungwon Kim, Kevin J Shih, Rohan Badlani, Joao Felipe Santos, Evelina Bhakturina, Mikyas Desta, Rafael Valle, Sungroh Yoon, Bryan Catanzaro

Affiliations: NVIDIA

Status : Added newer samples with better prosody and pronunciation. Check out samples folder. LJSpeech pretrained ckpt - GDrive Link

Credits

Dry run

Quick run in Google Colab

Instructions to run

Architecture details

TODOs, features and update notes

Status : Added newer samples with better prosody and pronunciation. Check out `samples` folder. LJSpeech pretrained ckpt - GDrive Link