Adding alginment information to readme

.pre-commit-config.yaml

@@ -1,9 +1,9 @@
 default_language_version:
-  python: python3.10
+  python: python3.11
 
 repos:
   - repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v4.5.0
+    rev: v4.4.0
     hooks:
       # list of supported hooks: https://pre-commit.com/hooks.html
       - id: trailing-whitespace
@@ -18,28 +18,28 @@ repos:
 
   # python code formatting
   - repo: https://github.com/psf/black
-    rev: 23.12.1
+    rev: 23.9.1
     hooks:
       - id: black
         args: [--line-length, "120"]
 
   # python import sorting
   - repo: https://github.com/PyCQA/isort
-    rev: 5.13.2
+    rev: 5.12.0
     hooks:
       - id: isort
         args: ["--profile", "black", "--filter-files"]
 
   # python upgrading syntax to newer version
   - repo: https://github.com/asottile/pyupgrade
-    rev: v3.15.0
+    rev: v3.14.0
     hooks:
       - id: pyupgrade
         args: [--py38-plus]
 
   # python check (PEP8), programming errors and code complexity
   - repo: https://github.com/PyCQA/flake8
-    rev: 7.0.0
+    rev: 6.1.0
     hooks:
       - id: flake8
         args:
@@ -54,6 +54,6 @@ repos:
 
   # pylint
   - repo: https://github.com/pycqa/pylint
-    rev: v3.0.3
+    rev: v3.0.0
     hooks:
     -   id: pylint

README.md

@@ -17,7 +17,7 @@
 
 </div>
 
-> This is the official code implementation of 🍵 Matcha-TTS [ICASSP 2024].
+> This is the official code implementation of 🍵 Matcha-TTS.
 
 We propose 🍵 Matcha-TTS, a new approach to non-autoregressive neural TTS, that uses [conditional flow matching](https://arxiv.org/abs/2210.02747) (similar to [rectified flows](https://arxiv.org/abs/2209.03003)) to speed up ODE-based speech synthesis. Our method:
 
@@ -252,6 +252,43 @@ python3 -m matcha.onnx.infer model.onnx --text "hey" --output-dir ./outputs --vo
 
 This will write `.wav` audio files to the output directory.
 
+## Extract phoneme alignments from Matcha-TTS
+
+If the dataset is structured as
+
+```bash
+data/
+└── LJSpeech-1.1
+    ├── metadata.csv
+    ├── README
+    ├── test.txt
+    ├── train.txt
+    ├── val.txt
+    └── wavs
+```
+Then you can extract the phoneme level alignments from a Trained Matcha-TTS model using:
+```bash
+python  matcha/utils/get_durations_from_trained_model.py -i dataset_yaml -c <checkpoint>
+```
+Example:
+```bash
+python  matcha/utils/get_durations_from_trained_model.py -i ljspeech.yaml -c matcha_ljspeech.ckpt
+```
+or simply:
+```bash
+matcha-tts-get-durations -i ljspeech.yaml -c matcha_ljspeech.ckpt
+```
+---
+## Train using extracted alignments
+
+In the datasetconfig turn on load duration.
+Example: `ljspeech.yaml`
+```
+load_durations: True
+```
+or see an examples in configs/experiment/ljspeech_from_durations.yaml
+
+
 ## Citation information
 
 If you use our code or otherwise find this work useful, please cite our paper:

configs/data/joe_spont_only.yaml

@@ -1,10 +0,0 @@
-defaults:
-  - ljspeech
-  - _self_
-
-name: joe_spont_only
-train_filelist_path: data/filelists/joe_spontonly_train.txt
-valid_filelist_path: data/filelists/joe_spontonly_val.txt
-data_statistics:
-  mel_mean: -5.882903
-  mel_std: 2.458284

configs/data/ljspeech.yaml

@@ -1,7 +1,7 @@
 _target_: matcha.data.text_mel_datamodule.TextMelDataModule
 name: ljspeech
-train_filelist_path: data/filelists/ljs_audio_text_train_filelist.txt
+train_filelist_path: data/LJSpeech-1.1/train.txt
-valid_filelist_path: data/filelists/ljs_audio_text_val_filelist.txt
+valid_filelist_path: data/LJSpeech-1.1/val.txt
 batch_size: 32
 num_workers: 20
 pin_memory: True
@@ -19,3 +19,4 @@ data_statistics:  # Computed for ljspeech dataset
   mel_mean: -5.536622
   mel_std: 2.116101
 seed: ${seed}
+load_durations: false

configs/data/ryan.yaml

@@ -1,10 +0,0 @@
-defaults:
-  - ljspeech
-  - _self_
-
-name: ryan
-train_filelist_path: data/filelists/ryan_train.csv
-valid_filelist_path: data/filelists/ryan_val.csv
-data_statistics:
-  mel_mean: -4.715779
-  mel_std: 2.124502

configs/data/tsg2.yaml

@@ -1,10 +0,0 @@
-defaults:
-  - ljspeech
-  - _self_
-
-name: tsg2
-train_filelist_path: data/filelists/cormac_train.txt
-valid_filelist_path: data/filelists/cormac_val.txt
-data_statistics:
-  mel_mean: -5.536622
-  mel_std: 2.116101

configs/experiment/joe_det_dur.yaml

@@ -1,14 +0,0 @@
-# @package _global_
-
-# to execute this experiment run:
-# python train.py experiment=multispeaker
-
-defaults:
-  - override /data: joe_spont_only.yaml
-
-# all parameters below will be merged with parameters from default configurations set above
-# this allows you to overwrite only specified parameters
-
-tags: ["joe"]
-
-run_name: joe_det_dur

configs/experiment/joe_stoc_dur.yaml

@@ -1,20 +0,0 @@
-# @package _global_
-
-# to execute this experiment run:
-# python train.py experiment=multispeaker
-
-defaults:
-  - override /data: joe_spont_only.yaml
-  - override /model/duration_predictor: flow_matching.yaml
-
-# all parameters below will be merged with parameters from default configurations set above
-# this allows you to overwrite only specified parameters
-
-tags: ["joe"]
-
-
-run_name: joe_stoc_dur
-
-model:
-  duration_predictor:
-    p_dropout: 0.2

configs/experiment/ljspeech_from_durations.yaml

@@ -5,12 +5,15 @@
 
 defaults:
   - override /data: ljspeech.yaml
-  - override /model/duration_predictor: flow_matching.yaml
 
 # all parameters below will be merged with parameters from default configurations set above
 # this allows you to overwrite only specified parameters
 
 tags: ["ljspeech"]
 
-
 run_name: ljspeech
+
+
+data:
+  load_durations: True
+  batch_size: 64

configs/experiment/ryan_det_dur.yaml

@@ -1,18 +0,0 @@
-# @package _global_
-
-# to execute this experiment run:
-# python train.py experiment=multispeaker
-
-defaults:
-  - override /data: ryan.yaml
-
-# all parameters below will be merged with parameters from default configurations set above
-# this allows you to overwrite only specified parameters
-
-tags: ["ryan"]
-
-run_name: ryan_det_dur
-
-
-trainer:
-  max_epochs: 3000

configs/experiment/ryan_stoc_dur.yaml

@@ -1,24 +0,0 @@
-# @package _global_
-
-# to execute this experiment run:
-# python train.py experiment=multispeaker
-
-defaults:
-  - override /data: ryan.yaml
-  - override /model/duration_predictor: flow_matching.yaml
-
-# all parameters below will be merged with parameters from default configurations set above
-# this allows you to overwrite only specified parameters
-
-tags: ["ryan"]
-
-
-run_name: ryan_stoc_dur
-
-model:
-  duration_predictor:
-    p_dropout: 0.2
-
-
-trainer:
-  max_epochs: 3000

configs/experiment/tsg2_det_dur.yaml

@@ -1,14 +0,0 @@
-# @package _global_
-
-# to execute this experiment run:
-# python train.py experiment=multispeaker
-
-defaults:
-  - override /data: tsg2.yaml
-
-# all parameters below will be merged with parameters from default configurations set above
-# this allows you to overwrite only specified parameters
-
-tags: ["tsg2"]
-
-run_name: tsg2_det_dur

configs/experiment/tsg2_stoc_dur.yaml

@@ -1,20 +0,0 @@
-# @package _global_
-
-# to execute this experiment run:
-# python train.py experiment=multispeaker
-
-defaults:
-  - override /data: tsg2.yaml
-  - override /model/duration_predictor: flow_matching.yaml
-
-# all parameters below will be merged with parameters from default configurations set above
-# this allows you to overwrite only specified parameters
-
-tags: ["tsg2"]
-
-
-run_name: tsg2_stoc_dur
-
-model:
-  duration_predictor:
-    p_dropout: 0.5

configs/model/duration_predictor/deterministic.yaml

@@ -1,7 +0,0 @@
-name: deterministic
-n_spks: ${model.n_spks}
-spk_emb_dim: ${model.spk_emb_dim}
-filter_channels: 256
-kernel_size: 3
-n_channels: ${model.encoder.encoder_params.n_channels}
-p_dropout: ${model.encoder.encoder_params.p_dropout}

configs/model/duration_predictor/flow_matching.yaml

@@ -1,7 +0,0 @@
-defaults:
-  - deterministic.yaml
-  - _self_
-
-sigma_min: 1e-4
-n_steps: 10
-name: flow_matching

configs/model/encoder/default.yaml

@@ -3,8 +3,16 @@ encoder_params:
   n_feats: ${model.n_feats}
   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
+
+duration_predictor_params:
+  filter_channels_dp: ${model.encoder.encoder_params.filter_channels_dp}
+  kernel_size: 3
+  p_dropout: ${model.encoder.encoder_params.p_dropout}

configs/model/matcha.yaml

@@ -1,7 +1,6 @@
 defaults:
   - _self_
   - encoder: default.yaml
-  - duration_predictor: deterministic.yaml
   - decoder: default.yaml
   - cfm: default.yaml
   - optimizer: adam.yaml
@@ -14,3 +13,4 @@ n_feats: 80
 data_statistics: ${data.data_statistics}
 out_size: null # Must be divisible by 4
 prior_loss: true
+use_precomputed_durations: ${data.load_durations}

matcha/VERSION

@@ -1 +1 @@
-0.0.5.1
+0.0.6.0

matcha/cli.py

@@ -48,7 +48,7 @@ def plot_spectrogram_to_numpy(spectrogram, filename):
 def process_text(i: int, text: str, device: torch.device):
     print(f"[{i}] - Input text: {text}")
     x = torch.tensor(
-        intersperse(text_to_sequence(text, ["english_cleaners2"]), 0),
+        intersperse(text_to_sequence(text, ["english_cleaners2"])[0], 0),
         dtype=torch.long,
         device=device,
     )[None]
@@ -227,7 +227,7 @@ def cli():
     parser.add_argument(
         "--vocoder",
         type=str,
-        default="hifigan_univ_v1",
+        default=None,
         help="Vocoder to use (default: will use the one suggested with the pretrained model))",
         choices=VOCODER_URLS.keys(),
     )
@@ -326,12 +326,13 @@ def batched_synthesis(args, device, model, vocoder, denoiser, texts, spk):
     for i, batch in enumerate(dataloader):
         i = i + 1
         start_t = dt.datetime.now()
+        b = batch["x"].shape[0]
         output = model.synthesise(
             batch["x"].to(device),
             batch["x_lengths"].to(device),
             n_timesteps=args.steps,
             temperature=args.temperature,
-            spks=spk,
+            spks=spk.expand(b) if spk is not None else spk,
             length_scale=args.speaking_rate,
         )
 

matcha/data/text_mel_datamodule.py

@@ -1,6 +1,8 @@
 import random
+from pathlib import Path
 from typing import Any, Dict, Optional
 
+import numpy as np
 import torch
 import torchaudio as ta
 from lightning import LightningDataModule
@@ -39,6 +41,7 @@ class TextMelDataModule(LightningDataModule):
         f_max,
         data_statistics,
         seed,
+        load_durations,
     ):
         super().__init__()
 
@@ -68,6 +71,7 @@ class TextMelDataModule(LightningDataModule):
             self.hparams.f_max,
             self.hparams.data_statistics,
             self.hparams.seed,
+            self.hparams.load_durations,
         )
         self.validset = TextMelDataset(  # pylint: disable=attribute-defined-outside-init
             self.hparams.valid_filelist_path,
@@ -83,6 +87,7 @@ class TextMelDataModule(LightningDataModule):
             self.hparams.f_max,
             self.hparams.data_statistics,
             self.hparams.seed,
+            self.hparams.load_durations,
         )
 
     def train_dataloader(self):
@@ -134,6 +139,7 @@ class TextMelDataset(torch.utils.data.Dataset):
         f_max=8000,
         data_parameters=None,
         seed=None,
+        load_durations=False,
     ):
         self.filepaths_and_text = parse_filelist(filelist_path)
         self.n_spks = n_spks
@@ -146,6 +152,8 @@ class TextMelDataset(torch.utils.data.Dataset):
         self.win_length = win_length
         self.f_min = f_min
         self.f_max = f_max
+        self.load_durations = load_durations
+
         if data_parameters is not None:
             self.data_parameters = data_parameters
         else:
@@ -167,7 +175,26 @@ class TextMelDataset(torch.utils.data.Dataset):
         text, cleaned_text = self.get_text(text, add_blank=self.add_blank)
         mel = self.get_mel(filepath)
 
-        return {"x": text, "y": mel, "spk": spk, "filepath": filepath, "x_text": cleaned_text}
+        durations = self.get_durations(filepath, text) if self.load_durations else None
+
+        return {"x": text, "y": mel, "spk": spk, "filepath": filepath, "x_text": cleaned_text, "durations": durations}
+
+    def get_durations(self, filepath, text):
+        filepath = Path(filepath)
+        data_dir, name = filepath.parent.parent, filepath.stem
+
+        try:
+            dur_loc = data_dir / "durations" / f"{name}.npy"
+            durs = torch.from_numpy(np.load(dur_loc).astype(int))
+
+        except FileNotFoundError as e:
+            raise FileNotFoundError(
+                f"Tried loading the durations but durations didn't exist at {dur_loc}, make sure you've generate the durations first using: python matcha/utils/get_durations_from_trained_model.py \n"
+            ) from e
+
+        assert len(durs) == len(text), f"Length of durations {len(durs)} and text {len(text)} do not match"
+
+        return durs
 
     def get_mel(self, filepath):
         audio, sr = ta.load(filepath)
@@ -207,13 +234,15 @@ class TextMelBatchCollate:
 
     def __call__(self, batch):
         B = len(batch)
-        y_max_length = max([item["y"].shape[-1] for item in batch])  # pylint: disable=consider-using-generator
+        y_max_length = max([item["y"].shape[-1] for item in batch])
         y_max_length = fix_len_compatibility(y_max_length)
-        x_max_length = max([item["x"].shape[-1] for item in batch])  # pylint: disable=consider-using-generator
+        x_max_length = max([item["x"].shape[-1] for item in batch])
         n_feats = batch[0]["y"].shape[-2]
 
         y = torch.zeros((B, n_feats, y_max_length), dtype=torch.float32)
         x = torch.zeros((B, x_max_length), dtype=torch.long)
+        durations = torch.zeros((B, x_max_length), dtype=torch.long)
+
         y_lengths, x_lengths = [], []
         spks = []
         filepaths, x_texts = [], []
@@ -226,6 +255,8 @@ class TextMelBatchCollate:
             spks.append(item["spk"])
             filepaths.append(item["filepath"])
             x_texts.append(item["x_text"])
+            if item["durations"] is not None:
+                durations[i, : item["durations"].shape[-1]] = item["durations"]
 
         y_lengths = torch.tensor(y_lengths, dtype=torch.long)
         x_lengths = torch.tensor(x_lengths, dtype=torch.long)
@@ -239,4 +270,5 @@ class TextMelBatchCollate:
             "spks": spks,
             "filepaths": filepaths,
             "x_texts": x_texts,
+            "durations": durations if not torch.eq(durations, 0).all() else None,
         }

matcha/models/baselightningmodule.py

@@ -65,6 +65,7 @@ class BaseLightningClass(LightningModule, ABC):
             y_lengths=y_lengths,
             spks=spks,
             out_size=self.out_size,
+            durations=batch["durations"],
         )
         return {
             "dur_loss": dur_loss,

matcha/models/components/duration_predictors.py

@@ -1,448 +0,0 @@
-import math
-
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-from einops import pack
-
-from matcha.models.components.decoder import SinusoidalPosEmb, TimestepEmbedding
-from matcha.models.components.text_encoder import LayerNorm
-
-# Define available networks
-
-
-class DurationPredictorNetwork(nn.Module):
-    def __init__(self, in_channels, filter_channels, kernel_size, p_dropout):
-        super().__init__()
-
-        self.in_channels = in_channels
-        self.filter_channels = filter_channels
-        self.p_dropout = p_dropout
-
-        self.drop = torch.nn.Dropout(p_dropout)
-        self.conv_1 = torch.nn.Conv1d(in_channels, filter_channels, kernel_size, padding=kernel_size // 2)
-        self.norm_1 = LayerNorm(filter_channels)
-        self.conv_2 = torch.nn.Conv1d(filter_channels, filter_channels, kernel_size, padding=kernel_size // 2)
-        self.norm_2 = LayerNorm(filter_channels)
-        self.proj = torch.nn.Conv1d(filter_channels, 1, 1)
-
-    def forward(self, x, x_mask):
-        x = self.conv_1(x * x_mask)
-        x = torch.relu(x)
-        x = self.norm_1(x)
-        x = self.drop(x)
-        x = self.conv_2(x * x_mask)
-        x = torch.relu(x)
-        x = self.norm_2(x)
-        x = self.drop(x)
-        x = self.proj(x * x_mask)
-        return x * x_mask
-
-
-class DurationPredictorNetworkWithTimeStep(nn.Module):
-    """Similar architecture but with a time embedding support"""
-
-    def __init__(self, in_channels, filter_channels, kernel_size, p_dropout):
-        super().__init__()
-        self.in_channels = in_channels
-        self.filter_channels = filter_channels
-        self.p_dropout = p_dropout
-
-        self.time_embeddings = SinusoidalPosEmb(filter_channels)
-        self.time_mlp = TimestepEmbedding(
-            in_channels=filter_channels,
-            time_embed_dim=filter_channels,
-            act_fn="silu",
-        )
-
-        self.drop = torch.nn.Dropout(p_dropout)
-        self.conv_1 = torch.nn.Conv1d(in_channels, filter_channels, kernel_size, padding=kernel_size // 2)
-        self.norm_1 = LayerNorm(filter_channels)
-        self.conv_2 = torch.nn.Conv1d(filter_channels, filter_channels, kernel_size, padding=kernel_size // 2)
-        self.norm_2 = LayerNorm(filter_channels)
-        self.proj = torch.nn.Conv1d(filter_channels, 1, 1)
-
-    def forward(self, x, x_mask, enc_outputs, t):
-        t = self.time_embeddings(t)
-        t = self.time_mlp(t).unsqueeze(-1)
-
-        x = pack([x, enc_outputs], "b * t")[0]
-
-        x = self.conv_1(x * x_mask)
-        x = torch.relu(x)
-        x = x + t
-        x = self.norm_1(x)
-        x = self.drop(x)
-        x = self.conv_2(x * x_mask)
-        x = torch.relu(x)
-        x = x + t
-        x = self.norm_2(x)
-        x = self.drop(x)
-        x = self.proj(x * x_mask)
-        return x * x_mask
-
-
-# Define available methods to compute loss
-
-# Simple MSE deterministic
-
-
-class DeterministicDurationPredictor(nn.Module):
-    def __init__(self, params):
-        super().__init__()
-        self.estimator = DurationPredictorNetwork(
-            params.n_channels + (params.spk_emb_dim if params.n_spks > 1 else 0),
-            params.filter_channels,
-            params.kernel_size,
-            params.p_dropout,
-        )
-
-    @torch.inference_mode()
-    def forward(self, x, x_mask):
-        return self.estimator(x, x_mask)
-
-    def compute_loss(self, durations, enc_outputs, x_mask):
-        return F.mse_loss(self.estimator(enc_outputs, x_mask), durations, reduction="sum") / torch.sum(x_mask)
-
-
-# Flow Matching duration predictor
-
-
-class FlowMatchingDurationPrediction(nn.Module):
-    def __init__(self, params) -> None:
-        super().__init__()
-
-        self.estimator = DurationPredictorNetworkWithTimeStep(
-            1
-            + params.n_channels
-            + (
-                params.spk_emb_dim if params.n_spks > 1 else 0
-            ),  # 1 for the durations and n_channels for encoder outputs
-            params.filter_channels,
-            params.kernel_size,
-            params.p_dropout,
-        )
-        self.sigma_min = params.sigma_min
-        self.n_steps = params.n_steps
-
-    @torch.inference_mode()
-    def forward(self, enc_outputs, mask, n_timesteps=500, temperature=1):
-        """Forward diffusion
-
-        Args:
-            mu (torch.Tensor): output of encoder
-                shape: (batch_size, n_feats, mel_timesteps)
-            mask (torch.Tensor): output_mask
-                shape: (batch_size, 1, mel_timesteps)
-            n_timesteps (int): number of diffusion steps
-            temperature (float, optional): temperature for scaling noise. Defaults to 1.0.
-            spks (torch.Tensor, optional): speaker ids. Defaults to None.
-                shape: (batch_size, spk_emb_dim)
-            cond: Not used but kept for future purposes
-
-        Returns:
-            sample: generated mel-spectrogram
-                shape: (batch_size, n_feats, mel_timesteps)
-        """
-        if n_timesteps is None:
-            n_timesteps = self.n_steps
-
-        b, _, t = enc_outputs.shape
-        z = torch.randn((b, 1, t), device=enc_outputs.device, dtype=enc_outputs.dtype) * temperature
-        t_span = torch.linspace(0, 1, n_timesteps + 1, device=enc_outputs.device)
-        return self.solve_euler(z, t_span=t_span, enc_outputs=enc_outputs, mask=mask)
-
-    def solve_euler(self, x, t_span, enc_outputs, mask):
-        """
-        Fixed euler solver for ODEs.
-        Args:
-            x (torch.Tensor): random noise
-            t_span (torch.Tensor): n_timesteps interpolated
-                shape: (n_timesteps + 1,)
-            mu (torch.Tensor): output of encoder
-                shape: (batch_size, n_feats, mel_timesteps)
-            mask (torch.Tensor): output_mask
-                shape: (batch_size, 1, mel_timesteps)
-            spks (torch.Tensor, optional): speaker ids. Defaults to None.
-                shape: (batch_size, spk_emb_dim)
-        """
-        t, _, dt = t_span[0], t_span[-1], t_span[1] - t_span[0]
-
-        # I am storing this because I can later plot it by putting a debugger here and saving it to a file
-        # Or in future might add like a return_all_steps flag
-        sol = []
-
-        for step in range(1, len(t_span)):
-            dphi_dt = self.estimator(x, mask, enc_outputs, t)
-
-            x = x + dt * dphi_dt
-            t = t + dt
-            sol.append(x)
-            if step < len(t_span) - 1:
-                dt = t_span[step + 1] - t
-
-        return sol[-1]
-
-    def compute_loss(self, x1, enc_outputs, mask):
-        """Computes diffusion loss
-
-        Args:
-            x1 (torch.Tensor): Target
-                shape: (batch_size, n_feats, mel_timesteps)
-            mask (torch.Tensor): target mask
-                shape: (batch_size, 1, mel_timesteps)
-            mu (torch.Tensor): output of encoder
-                shape: (batch_size, n_feats, mel_timesteps)
-            spks (torch.Tensor, optional): speaker embedding. Defaults to None.
-                shape: (batch_size, spk_emb_dim)
-
-        Returns:
-            loss: conditional flow matching loss
-            y: conditional flow
-                shape: (batch_size, n_feats, mel_timesteps)
-        """
-        enc_outputs = enc_outputs.detach()  # don't update encoder from the duration predictor
-        b, _, t = enc_outputs.shape
-
-        # random timestep
-        t = torch.rand([b, 1, 1], device=enc_outputs.device, dtype=enc_outputs.dtype)
-        # sample noise p(x_0)
-        z = torch.randn_like(x1)
-
-        y = (1 - (1 - self.sigma_min) * t) * z + t * x1
-        u = x1 - (1 - self.sigma_min) * z
-
-        loss = F.mse_loss(self.estimator(y, mask, enc_outputs, t.squeeze()), u, reduction="sum") / (
-            torch.sum(mask) * u.shape[1]
-        )
-        return loss
-
-
-# VITS discrete normalising flow based duration predictor
-
-
-class Log(nn.Module):
-    def forward(self, x, x_mask, reverse=False, **kwargs):
-        if not reverse:
-            y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask
-            logdet = torch.sum(-y, [1, 2])
-            return y, logdet
-        else:
-            x = torch.exp(x) * x_mask
-            return x
-
-
-class ElementwiseAffine(nn.Module):
-    def __init__(self, channels):
-        super().__init__()
-        self.channels = channels
-        self.m = nn.Parameter(torch.zeros(channels, 1))
-        self.logs = nn.Parameter(torch.zeros(channels, 1))
-
-    def forward(self, x, x_mask, reverse=False, **kwargs):
-        if not reverse:
-            y = self.m + torch.exp(self.logs) * x
-            y = y * x_mask
-            logdet = torch.sum(self.logs * x_mask, [1, 2])
-            return y, logdet
-        else:
-            x = (x - self.m) * torch.exp(-self.logs) * x_mask
-            return x
-
-
-class DDSConv(nn.Module):
-    """
-    Dialted and Depth-Separable Convolution
-    """
-
-    def __init__(self, channels, kernel_size, n_layers, p_dropout=0.0):
-        super().__init__()
-        self.channels = channels
-        self.kernel_size = kernel_size
-        self.n_layers = n_layers
-        self.p_dropout = p_dropout
-
-        self.drop = nn.Dropout(p_dropout)
-        self.convs_sep = nn.ModuleList()
-        self.convs_1x1 = nn.ModuleList()
-        self.norms_1 = nn.ModuleList()
-        self.norms_2 = nn.ModuleList()
-        for i in range(n_layers):
-            dilation = kernel_size**i
-            padding = (kernel_size * dilation - dilation) // 2
-            self.convs_sep.append(
-                nn.Conv1d(channels, channels, kernel_size, groups=channels, dilation=dilation, padding=padding)
-            )
-            self.convs_1x1.append(nn.Conv1d(channels, channels, 1))
-            self.norms_1.append(LayerNorm(channels))
-            self.norms_2.append(LayerNorm(channels))
-
-    def forward(self, x, x_mask, g=None):
-        if g is not None:
-            x = x + g
-        for i in range(self.n_layers):
-            y = self.convs_sep[i](x * x_mask)
-            y = self.norms_1[i](y)
-            y = F.gelu(y)
-            y = self.convs_1x1[i](y)
-            y = self.norms_2[i](y)
-            y = F.gelu(y)
-            y = self.drop(y)
-            x = x + y
-        return x * x_mask
-
-
-class ConvFlow(nn.Module):
-    def __init__(self, in_channels, filter_channels, kernel_size, n_layers, num_bins=10, tail_bound=5.0):
-        super().__init__()
-        self.in_channels = in_channels
-        self.filter_channels = filter_channels
-        self.kernel_size = kernel_size
-        self.n_layers = n_layers
-        self.num_bins = num_bins
-        self.tail_bound = tail_bound
-        self.half_channels = in_channels // 2
-
-        self.pre = nn.Conv1d(self.half_channels, filter_channels, 1)
-        self.convs = DDSConv(filter_channels, kernel_size, n_layers, p_dropout=0.0)
-        self.proj = nn.Conv1d(filter_channels, self.half_channels * (num_bins * 3 - 1), 1)
-        self.proj.weight.data.zero_()
-        self.proj.bias.data.zero_()
-
-    def forward(self, x, x_mask, g=None, reverse=False):
-        x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
-        h = self.pre(x0)
-        h = self.convs(h, x_mask, g=g)
-        h = self.proj(h) * x_mask
-
-        b, c, t = x0.shape
-        h = h.reshape(b, c, -1, t).permute(0, 1, 3, 2)  # [b, cx?, t] -> [b, c, t, ?]
-
-        unnormalized_widths = h[..., : self.num_bins] / math.sqrt(self.filter_channels)
-        unnormalized_heights = h[..., self.num_bins : 2 * self.num_bins] / math.sqrt(self.filter_channels)
-        unnormalized_derivatives = h[..., 2 * self.num_bins :]
-
-        x1, logabsdet = piecewise_rational_quadratic_transform(
-            x1,
-            unnormalized_widths,
-            unnormalized_heights,
-            unnormalized_derivatives,
-            inverse=reverse,
-            tails="linear",
-            tail_bound=self.tail_bound,
-        )
-
-        x = torch.cat([x0, x1], 1) * x_mask
-        logdet = torch.sum(logabsdet * x_mask, [1, 2])
-        if not reverse:
-            return x, logdet
-        else:
-            return x
-
-
-class StochasticDurationPredictor(nn.Module):
-    def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, n_flows=4, gin_channels=0):
-        super().__init__()
-        filter_channels = in_channels  # it needs to be removed from future version.
-        self.in_channels = in_channels
-        self.filter_channels = filter_channels
-        self.kernel_size = kernel_size
-        self.p_dropout = p_dropout
-        self.n_flows = n_flows
-        self.gin_channels = gin_channels
-
-        self.log_flow = Log()
-        self.flows = nn.ModuleList()
-        self.flows.append(ElementwiseAffine(2))
-        for i in range(n_flows):
-            self.flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))
-            self.flows.append(modules.Flip())
-
-        self.post_pre = nn.Conv1d(1, filter_channels, 1)
-        self.post_proj = nn.Conv1d(filter_channels, filter_channels, 1)
-        self.post_convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)
-        self.post_flows = nn.ModuleList()
-        self.post_flows.append(modules.ElementwiseAffine(2))
-        for i in range(4):
-            self.post_flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))
-            self.post_flows.append(modules.Flip())
-
-        self.pre = nn.Conv1d(in_channels, filter_channels, 1)
-        self.proj = nn.Conv1d(filter_channels, filter_channels, 1)
-        self.convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)
-        if gin_channels != 0:
-            self.cond = nn.Conv1d(gin_channels, filter_channels, 1)
-
-    def forward(self, x, x_mask, w=None, g=None, reverse=False, noise_scale=1.0):
-        x = torch.detach(x)
-        x = self.pre(x)
-        if g is not None:
-            g = torch.detach(g)
-            x = x + self.cond(g)
-        x = self.convs(x, x_mask)
-        x = self.proj(x) * x_mask
-
-        if not reverse:
-            flows = self.flows
-            assert w is not None
-
-            logdet_tot_q = 0
-            h_w = self.post_pre(w)
-            h_w = self.post_convs(h_w, x_mask)
-            h_w = self.post_proj(h_w) * x_mask
-            e_q = torch.randn(w.size(0), 2, w.size(2)).to(device=x.device, dtype=x.dtype) * x_mask
-            z_q = e_q
-            for flow in self.post_flows:
-                z_q, logdet_q = flow(z_q, x_mask, g=(x + h_w))
-                logdet_tot_q += logdet_q
-            z_u, z1 = torch.split(z_q, [1, 1], 1)
-            u = torch.sigmoid(z_u) * x_mask
-            z0 = (w - u) * x_mask
-            logdet_tot_q += torch.sum((F.logsigmoid(z_u) + F.logsigmoid(-z_u)) * x_mask, [1, 2])
-            logq = torch.sum(-0.5 * (math.log(2 * math.pi) + (e_q**2)) * x_mask, [1, 2]) - logdet_tot_q
-
-            logdet_tot = 0
-            z0, logdet = self.log_flow(z0, x_mask)
-            logdet_tot += logdet
-            z = torch.cat([z0, z1], 1)
-            for flow in flows:
-                z, logdet = flow(z, x_mask, g=x, reverse=reverse)
-                logdet_tot = logdet_tot + logdet
-            nll = torch.sum(0.5 * (math.log(2 * math.pi) + (z**2)) * x_mask, [1, 2]) - logdet_tot
-            return nll + logq  # [b]
-        else:
-            flows = list(reversed(self.flows))
-            flows = flows[:-2] + [flows[-1]]  # remove a useless vflow
-            z = torch.randn(x.size(0), 2, x.size(2)).to(device=x.device, dtype=x.dtype) * noise_scale
-            for flow in flows:
-                z = flow(z, x_mask, g=x, reverse=reverse)
-            z0, z1 = torch.split(z, [1, 1], 1)
-            logw = z0
-            return logw
-
-
-# Meta class to wrap all duration predictors
-
-
-class DP(nn.Module):
-    def __init__(self, params):
-        super().__init__()
-        self.name = params.name
-
-        if params.name == "deterministic":
-            self.dp = DeterministicDurationPredictor(
-                params,
-            )
-        elif params.name == "flow_matching":
-            self.dp = FlowMatchingDurationPrediction(
-                params,
-            )
-        else:
-            raise ValueError(f"Invalid duration predictor configuration: {params.name}")
-
-    @torch.inference_mode()
-    def forward(self, enc_outputs, mask):
-        return self.dp(enc_outputs, mask)
-
-    def compute_loss(self, durations, enc_outputs, mask):
-        return self.dp.compute_loss(durations, enc_outputs, mask)

matcha/models/components/text_encoder.py

@@ -67,6 +67,33 @@ class ConvReluNorm(nn.Module):
         return x * x_mask
 
 
+class DurationPredictor(nn.Module):
+    def __init__(self, in_channels, filter_channels, kernel_size, p_dropout):
+        super().__init__()
+        self.in_channels = in_channels
+        self.filter_channels = filter_channels
+        self.p_dropout = p_dropout
+
+        self.drop = torch.nn.Dropout(p_dropout)
+        self.conv_1 = torch.nn.Conv1d(in_channels, filter_channels, kernel_size, padding=kernel_size // 2)
+        self.norm_1 = LayerNorm(filter_channels)
+        self.conv_2 = torch.nn.Conv1d(filter_channels, filter_channels, kernel_size, padding=kernel_size // 2)
+        self.norm_2 = LayerNorm(filter_channels)
+        self.proj = torch.nn.Conv1d(filter_channels, 1, 1)
+
+    def forward(self, x, x_mask):
+        x = self.conv_1(x * x_mask)
+        x = torch.relu(x)
+        x = self.norm_1(x)
+        x = self.drop(x)
+        x = self.conv_2(x * x_mask)
+        x = torch.relu(x)
+        x = self.norm_2(x)
+        x = self.drop(x)
+        x = self.proj(x * x_mask)
+        return x * x_mask
+
+
 class RotaryPositionalEmbeddings(nn.Module):
     """
     ## RoPE module
@@ -303,6 +330,7 @@ class TextEncoder(nn.Module):
         self,
         encoder_type,
         encoder_params,
+        duration_predictor_params,
         n_vocab,
         n_spks=1,
         spk_emb_dim=128,
@@ -340,6 +368,12 @@ class TextEncoder(nn.Module):
         )
 
         self.proj_m = torch.nn.Conv1d(self.n_channels + (spk_emb_dim if n_spks > 1 else 0), self.n_feats, 1)
+        self.proj_w = DurationPredictor(
+            self.n_channels + (spk_emb_dim if n_spks > 1 else 0),
+            duration_predictor_params.filter_channels_dp,
+            duration_predictor_params.kernel_size,
+            duration_predictor_params.p_dropout,
+        )
 
     def forward(self, x, x_lengths, spks=None):
         """Run forward pass to the transformer based encoder and duration predictor
@@ -370,7 +404,7 @@ class TextEncoder(nn.Module):
         x = self.encoder(x, x_mask)
         mu = self.proj_m(x) * x_mask
 
-        # x_dp = torch.detach(x)
+        x_dp = torch.detach(x)
-        # logw = self.proj_w(x_dp, x_mask)
+        logw = self.proj_w(x_dp, x_mask)
 
-        return mu, x, x_mask
+        return mu, logw, x_mask

matcha/models/matcha_tts.py

@@ -4,14 +4,14 @@ import random
 
 import torch
 
-import matcha.utils.monotonic_align as monotonic_align  # pylint: disable=consider-using-from-import
+import matcha.utils.monotonic_align as monotonic_align
 from matcha import utils
 from matcha.models.baselightningmodule import BaseLightningClass
-from matcha.models.components.duration_predictors import DP
 from matcha.models.components.flow_matching import CFM
 from matcha.models.components.text_encoder import TextEncoder
 from matcha.utils.model import (
     denormalize,
+    duration_loss,
     fix_len_compatibility,
     generate_path,
     sequence_mask,
@@ -28,7 +28,6 @@ class MatchaTTS(BaseLightningClass):  # 🍵
         spk_emb_dim,
         n_feats,
         encoder,
-        duration_predictor,
         decoder,
         cfm,
         data_statistics,
@@ -36,6 +35,7 @@ class MatchaTTS(BaseLightningClass):  # 🍵
         optimizer=None,
         scheduler=None,
         prior_loss=True,
+        use_precomputed_durations=False,
     ):
         super().__init__()
 
@@ -47,6 +47,7 @@ class MatchaTTS(BaseLightningClass):  # 🍵
         self.n_feats = n_feats
         self.out_size = out_size
         self.prior_loss = prior_loss
+        self.use_precomputed_durations = use_precomputed_durations
 
         if n_spks > 1:
             self.spk_emb = torch.nn.Embedding(n_spks, spk_emb_dim)
@@ -54,13 +55,12 @@ class MatchaTTS(BaseLightningClass):  # 🍵
         self.encoder = TextEncoder(
             encoder.encoder_type,
             encoder.encoder_params,
+            encoder.duration_predictor_params,
             n_vocab,
             n_spks,
             spk_emb_dim,
         )
 
-        self.dp = DP(duration_predictor)
-
         self.decoder = CFM(
             in_channels=2 * encoder.encoder_params.n_feats,
             out_channel=encoder.encoder_params.n_feats,
@@ -114,15 +114,11 @@ class MatchaTTS(BaseLightningClass):  # 🍵
             # Get speaker embedding
             spks = self.spk_emb(spks.long())
 
-        # Get encoder_outputs `mu_x` and encoded text `enc_output`
+        # Get encoder_outputs `mu_x` and log-scaled token durations `logw`
-        mu_x, enc_output, x_mask = self.encoder(x, x_lengths, spks)
+        mu_x, logw, x_mask = self.encoder(x, x_lengths, spks)
-
-        # Get log-scaled token durations `logw`
-        logw = self.dp(enc_output, x_mask)
 
         w = torch.exp(logw) * x_mask
         w_ceil = torch.ceil(w) * length_scale
-        # print(w_ceil)
         y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long()
         y_max_length = y_lengths.max()
         y_max_length_ = fix_len_compatibility(y_max_length)
@@ -153,7 +149,7 @@ class MatchaTTS(BaseLightningClass):  # 🍵
             "rtf": rtf,
         }
 
-    def forward(self, x, x_lengths, y, y_lengths, spks=None, out_size=None, cond=None):
+    def forward(self, x, x_lengths, y, y_lengths, spks=None, out_size=None, cond=None, durations=None):
         """
         Computes 3 losses:
             1. duration loss: loss between predicted token durations and those extracted by Monotinic Alignment Search (MAS).
@@ -179,27 +175,31 @@ class MatchaTTS(BaseLightningClass):  # 🍵
             spks = self.spk_emb(spks)
 
         # Get encoder_outputs `mu_x` and log-scaled token durations `logw`
-        mu_x, enc_output, x_mask = self.encoder(x, x_lengths, spks)
+        mu_x, logw, x_mask = self.encoder(x, x_lengths, spks)
         y_max_length = y.shape[-1]
 
         y_mask = sequence_mask(y_lengths, y_max_length).unsqueeze(1).to(x_mask)
         attn_mask = x_mask.unsqueeze(-1) * y_mask.unsqueeze(2)
 
-        # Use MAS to find most likely alignment `attn` between text and mel-spectrogram
+        if self.use_precomputed_durations:
-        with torch.no_grad():
+            attn = generate_path(durations.squeeze(1), attn_mask.squeeze(1))
-            const = -0.5 * math.log(2 * math.pi) * self.n_feats
+        else:
-            factor = -0.5 * torch.ones(mu_x.shape, dtype=mu_x.dtype, device=mu_x.device)
+            # Use MAS to find most likely alignment `attn` between text and mel-spectrogram
-            y_square = torch.matmul(factor.transpose(1, 2), y**2)
+            with torch.no_grad():
-            y_mu_double = torch.matmul(2.0 * (factor * mu_x).transpose(1, 2), y)
+                const = -0.5 * math.log(2 * math.pi) * self.n_feats
-            mu_square = torch.sum(factor * (mu_x**2), 1).unsqueeze(-1)
+                factor = -0.5 * torch.ones(mu_x.shape, dtype=mu_x.dtype, device=mu_x.device)
-            log_prior = y_square - y_mu_double + mu_square + const
+                y_square = torch.matmul(factor.transpose(1, 2), y**2)
+                y_mu_double = torch.matmul(2.0 * (factor * mu_x).transpose(1, 2), y)
+                mu_square = torch.sum(factor * (mu_x**2), 1).unsqueeze(-1)
+                log_prior = y_square - y_mu_double + mu_square + const
 
-            attn = monotonic_align.maximum_path(log_prior, attn_mask.squeeze(1))
+                attn = monotonic_align.maximum_path(log_prior, attn_mask.squeeze(1))
-            attn = attn.detach()
+                attn = attn.detach()  # b, t_text, T_mel
 
         # Compute loss between predicted log-scaled durations and those obtained from MAS
+        # refered to as prior loss in the paper
         logw_ = torch.log(1e-8 + torch.sum(attn.unsqueeze(1), -1)) * x_mask
-        dur_loss = self.dp.compute_loss(logw_, enc_output, x_mask)
+        dur_loss = duration_loss(logw, logw_, x_lengths)
 
         # Cut a small segment of mel-spectrogram in order to increase batch size
         #   - "Hack" taken from Grad-TTS, in case of Grad-TTS, we cannot train batch size 32 on a 24GB GPU without it

matcha/text/cleaners.py

@@ -15,7 +15,6 @@ import logging
 import re
 
 import phonemizer
-import piper_phonemize
 from unidecode import unidecode
 
 # To avoid excessive logging we set the log level of the phonemizer package to Critical
@@ -106,11 +105,17 @@ def english_cleaners2(text):
     return phonemes
 
 
-def english_cleaners_piper(text):
+# I am removing this due to incompatibility with several version of python
-    """Pipeline for English text, including abbreviation expansion. + punctuation + stress"""
+# However, if you want to use it, you can uncomment it
-    text = convert_to_ascii(text)
+# and install piper-phonemize with the following command:
-    text = lowercase(text)
+# pip install piper-phonemize
-    text = expand_abbreviations(text)
+
-    phonemes = "".join(piper_phonemize.phonemize_espeak(text=text, voice="en-US")[0])
+# import piper_phonemize
-    phonemes = collapse_whitespace(phonemes)
+# def english_cleaners_piper(text):
-    return phonemes
+#     """Pipeline for English text, including abbreviation expansion. + punctuation + stress"""
+#     text = convert_to_ascii(text)
+#     text = lowercase(text)
+#     text = expand_abbreviations(text)
+#     phonemes = "".join(piper_phonemize.phonemize_espeak(text=text, voice="en-US")[0])
+#     phonemes = collapse_whitespace(phonemes)
+#     return phonemes

matcha/utils/generate_data_statistics.py

@@ -94,6 +94,7 @@ def main():
         cfg["batch_size"] = args.batch_size
         cfg["train_filelist_path"] = str(os.path.join(root_path, cfg["train_filelist_path"]))
         cfg["valid_filelist_path"] = str(os.path.join(root_path, cfg["valid_filelist_path"]))
+        cfg["load_durations"] = False
 
     text_mel_datamodule = TextMelDataModule(**cfg)
     text_mel_datamodule.setup()

matcha/utils/get_durations_from_trained_model.py

@@ -86,7 +86,7 @@ def main():
         "-i",
         "--input-config",
         type=str,
-        default="vctk.yaml",
+        default="ljspeech.yaml",
         help="The name of the yaml config file under configs/data",
     )
 
@@ -140,11 +140,14 @@ def main():
         cfg["batch_size"] = args.batch_size
         cfg["train_filelist_path"] = str(os.path.join(root_path, cfg["train_filelist_path"]))
         cfg["valid_filelist_path"] = str(os.path.join(root_path, cfg["valid_filelist_path"]))
+        cfg["load_durations"] = False
 
     if args.output_folder is not None:
         output_folder = Path(args.output_folder)
     else:
-        output_folder = Path("data") / "processed_data" / cfg["name"] / "durations"
+        output_folder = Path(cfg["train_filelist_path"]).parent / "durations"
+
+    print(f"Output folder set to: {output_folder}")
 
     if os.path.exists(output_folder) and not args.force:
         print("Folder already exists. Use -f to force overwrite")

requirements.txt

@@ -38,8 +38,7 @@ conformer==0.3.2
 diffusers==0.25.0
 notebook
 ipywidgets
-gradio
+gradio==3.43.2
 gdown
 wget
 seaborn
-piper_phonemize

scripts/get_durations.sh

@@ -1,15 +0,0 @@
-#!/bin/bash
-
-echo "Starting script"
-
-echo "Getting LJ Speech durations"
-python matcha/utils/get_durations_from_trained_model.py -i ljspeech.yaml -c logs/train/lj_det/runs/2024-01-12_12-05-00/checkpoints/last.ckpt -f
-
-echo "Getting TSG2 durations"
-python matcha/utils/get_durations_from_trained_model.py -i tsg2.yaml -c logs/train/tsg2_det_dur/runs/2024-01-05_12-33-35/checkpoints/last.ckpt -f
-
-echo "Getting Joe Spont durations"
-python matcha/utils/get_durations_from_trained_model.py -i joe_spont_only.yaml -c logs/train/joe_det_dur/runs/2024-02-20_14-01-01/checkpoints/last.ckpt -f
-
-echo "Getting Ryan durations"
-python matcha/utils/get_durations_from_trained_model.py -i ryan.yaml -c logs/train/matcha_ryan_det/runs/2024-02-26_09-28-09/checkpoints/last.ckpt -f

scripts/transcribe.sh

@@ -1,7 +0,0 @@
-echo "Transcribing"
-
-whispertranscriber -i lj_det_output -o lj_det_output_transcriptions -f
-
-whispertranscriber -i lj_fm_output -o lj_fm_output_transcriptions -f
-wercompute -r dur_wer_computation/reference_transcripts/ -i lj_det_output_transcriptions
-wercompute -r dur_wer_computation/reference_transcripts/ -i lj_fm_output_transcriptions

scripts/wer_computer.sh

@@ -1,30 +0,0 @@
-#!/bin/bash
-# Run from root folder with: bash scripts/wer_computer.sh
-
-
-root_folder=${1:-"dur_wer_computation"}
-echo "Running WER computation for Duration predictors"
-cmd="wercompute -r ${root_folder}/reference_transcripts/ -i ${root_folder}/lj_fm_output_transcriptions/"
-# echo $cmd
-echo "LJ"
-echo "==================================="
-echo "Flow Matching"
-$cmd
-echo "-----------------------------------"
-
-echo "LJ Determinstic"
-cmd="wercompute -r ${root_folder}/reference_transcripts/ -i ${root_folder}/lj_det_output_transcriptions/"
-$cmd
-echo "-----------------------------------"
-
-echo "Cormac"
-echo "==================================="
-echo "Cormac Flow Matching"
-cmd="wercompute -r ${root_folder}/reference_transcripts/ -i ${root_folder}/fm_output_transcriptions/"
-$cmd
-echo "-----------------------------------"
-
-echo "Cormac Determinstic"
-cmd="wercompute -r ${root_folder}/reference_transcripts/ -i ${root_folder}/det_output_transcriptions/"
-$cmd
-echo "-----------------------------------"

setup.py

@@ -38,6 +38,7 @@ setup(
             "matcha-data-stats=matcha.utils.generate_data_statistics:main",
             "matcha-tts=matcha.cli:cli",
             "matcha-tts-app=matcha.app:main",
+            "matcha-tts-get-durations=matcha.utils.get_durations_from_trained_model:main",
         ]
     },
     ext_modules=cythonize(exts, language_level=3),