use CMVN

egs/ljspeech/TTS/local/compute_fbank_ljspeech.py

@@ -106,6 +106,8 @@ def compute_fbank_ljspeech(num_jobs: int):
         use_fft_mag=True,
         low_freq=0,
         high_freq=8000,
+        remove_dc_offset=False,
+        preemph_coeff=0,
         # should be identical to n_feats in ../matcha/train.py
         num_filters=80,
     )

egs/ljspeech/TTS/local/compute_fbank_statistics.py

@@ -0,0 +1,84 @@
+#!/usr/bin/env python3
+# Copyright    2024  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script compute the mean and std of the fbank features.
+"""
+
+import argparse
+import json
+import logging
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, load_manifest_lazy
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "manifest",
+        type=Path,
+        help="Path to the manifest file",
+    )
+
+    parser.add_argument(
+        "cmvn",
+        type=Path,
+        help="Path to the cmvn.json",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+
+    manifest = args.manifest
+    logging.info(
+        f"Computing fbank mean and std for {manifest} and saving to {args.cmvn}"
+    )
+
+    assert manifest.is_file(), f"{manifest} does not exist"
+    cut_set = load_manifest_lazy(manifest)
+    assert isinstance(cut_set, CutSet), type(cut_set)
+
+    feat_dim = cut_set[0].features.num_features
+    num_frames = 0
+    s = 0
+    sq = 0
+    for c in cut_set:
+        f = torch.from_numpy(c.load_features())
+        num_frames += f.shape[0]
+        s += f.sum()
+        sq += f.square().sum()
+
+    fbank_mean = s / (num_frames * feat_dim)
+    fbank_var = sq / (num_frames * feat_dim) - fbank_mean * fbank_mean
+    print("fbank var", fbank_var)
+    fbank_std = fbank_var.sqrt()
+    with open(args.cmvn, "w") as f:
+        json.dump({"fbank_mean": fbank_mean.item(), "fbank_std": fbank_std.item()}, f)
+        f.write("\n")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()

egs/ljspeech/TTS/matcha/inference.py

@@ -10,6 +10,7 @@ import soundfile as sf
 import torch
 from matcha.hifigan.config import v1
 from matcha.hifigan.denoiser import Denoiser
+from tokenizer import Tokenizer
 from matcha.hifigan.models import Generator as HiFiGAN
 from matcha.text import sequence_to_text, text_to_sequence
 from matcha.utils.utils import intersperse
@@ -28,7 +29,7 @@ def get_parser():
     parser.add_argument(
         "--epoch",
         type=int,
-        default=140,
+        default=1320,
         help="""It specifies the checkpoint to use for decoding.
         Note: Epoch counts from 1.
         """,
@@ -37,13 +38,19 @@ def get_parser():
     parser.add_argument(
         "--exp-dir",
         type=Path,
-        default="matcha/exp",
+        default="matcha/exp-fbank",
         help="""The experiment dir.
         It specifies the directory where all training related
         files, e.g., checkpoints, log, etc, are saved
         """,
     )
 
+    parser.add_argument(
+        "--tokens",
+        type=Path,
+        default="data/tokens.txt",
+    )
+
     return parser
 
 
@@ -71,19 +78,17 @@ def save_to_folder(filename: str, output: dict, folder: str):
     sf.write(folder / f"{filename}.wav", output["waveform"], 22050, "PCM_24")
 
 
-def process_text(text: str):
-    x = torch.tensor(
-        intersperse(text_to_sequence(text, ["english_cleaners2"])[0], 0),
-        dtype=torch.long,
-        device="cpu",
-    )[None]
+def process_text(text: str, tokenizer):
+    x = tokenizer.texts_to_token_ids([text], add_sos=True, add_eos=True)
+    x = torch.tensor(x, dtype=torch.long)
     x_lengths = torch.tensor([x.shape[-1]], dtype=torch.long, device="cpu")
-    x_phones = sequence_to_text(x.squeeze(0).tolist())
-    return {"x_orig": text, "x": x, "x_lengths": x_lengths, "x_phones": x_phones}
+    return {"x_orig": text, "x": x, "x_lengths": x_lengths}
 
 
-def synthesise(model, n_timesteps, text, length_scale, temperature, spks=None):
-    text_processed = process_text(text)
+def synthesise(
+    model, tokenizer, n_timesteps, text, length_scale, temperature, spks=None
+):
+    text_processed = process_text(text, tokenizer)
     start_t = dt.datetime.now()
     output = model.synthesise(
         text_processed["x"],
@@ -108,6 +113,11 @@ def main():
     params.update(vars(args))
     logging.info(params)
 
+    tokenizer = Tokenizer(params.tokens)
+    params.blank_id = tokenizer.pad_id
+    params.vocab_size = tokenizer.vocab_size
+    params.model_args.n_vocab = params.vocab_size
+
     logging.info("About to create model")
     model = get_model(params)
     load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
@@ -117,12 +127,13 @@ def main():
     denoiser = Denoiser(vocoder, mode="zeros")
 
     texts = [
-        "The Secret Service believed that it was very doubtful that any President would ride regularly in a vehicle with a fixed top, even though transparent.",
-        "Today as always, men fall into two groups: slaves and free men. Whoever does not have two-thirds of his day for himself, is a slave, whatever he may be: a statesman, a businessman, an official, or a scholar.",
+        "How are you doing, my friend",
+        #  "The Secret Service believed that it was very doubtful that any President would ride regularly in a vehicle with a fixed top, even though transparent.",
+        #  "Today as always, men fall into two groups: slaves and free men. Whoever does not have two-thirds of his day for himself, is a slave, whatever he may be: a statesman, a businessman, an official, or a scholar.",
     ]
 
     # Number of ODE Solver steps
-    n_timesteps = 2
+    n_timesteps = 3
 
     # Changes to the speaking rate
     length_scale = 1.0
@@ -135,6 +146,7 @@ def main():
     for i, text in enumerate(tqdm(texts)):
         output = synthesise(
             model=model,
+            tokenizer=tokenizer,
             n_timesteps=n_timesteps,
             text=text,
             length_scale=length_scale,
@@ -154,7 +166,7 @@ def main():
         print(f"{'*' * 53}")
         print(f"Phonetised text - {i}")
         print(f"{'-' * 53}")
-        print(output["x_phones"])
+        print(output["x"])
         print(f"{'*' * 53}")
         print(f"RTF:\t\t{output['rtf']:.6f}")
         print(f"RTF Waveform:\t{rtf_w:.6f}")
@@ -162,7 +174,7 @@ def main():
         rtfs_w.append(rtf_w)
 
         # Save the generated waveform
-        save_to_folder(i, output, folder="./my-output")
+        save_to_folder(i, output, folder="./my-output-1320")
 
     print(f"Number of ODE steps: {n_timesteps}")
     print(f"Mean RTF:\t\t\t\t{np.mean(rtfs):.6f} ± {np.std(rtfs):.6f}")

egs/ljspeech/TTS/matcha/train.py

@@ -13,6 +13,7 @@ import torch
 import torch.multiprocessing as mp
 import torch.nn as nn
 from lhotse.utils import fix_random_seed
+from matcha.data.text_mel_datamodule import TextMelDataModule
 from matcha.models.matcha_tts import MatchaTTS
 from matcha.tokenizer import Tokenizer
 from matcha.utils.model import fix_len_compatibility
@@ -122,8 +123,11 @@ def get_parser():
 def get_data_statistics():
     return AttributeDict(
         {
-            "mel_mean": 0.0,
-            "mel_std": 1.0,
+            #  "mel_mean": -5.517028331756592, # matcha-tts
+            #  "mel_std": 2.0643954277038574,
+            # ours
+            "mel_mean": -1.168782114982605,
+            "mel_std": 1.9283572435379028,
         }
     )
 
@@ -134,7 +138,8 @@ def _get_data_params() -> AttributeDict:
             "name": "ljspeech",
             "train_filelist_path": "./filelists/ljs_audio_text_train_filelist.txt",
             "valid_filelist_path": "./filelists/ljs_audio_text_val_filelist.txt",
-            "num_workers": 3,
+            "batch_size": 64,
+            "num_workers": 1,
             "pin_memory": False,
             "cleaners": ["english_cleaners2"],
             "add_blank": True,
@@ -289,8 +294,17 @@ def load_checkpoint_if_available(
     return saved_params
 
 
-def prepare_input(batch: dict, tokenizer: Tokenizer, device: torch.device):
+def prepare_input(batch: dict, tokenizer: Tokenizer, device: torch.device, params):
     """Parse batch data"""
+    mel_mean = params.data_args.data_statistics.mel_mean
+    mel_std_inv = 1 / params.data_args.data_statistics.mel_std
+    for i in range(batch["features"].shape[0]):
+        n = batch["features_lens"][i]
+        batch["features"][i : i + 1, :n, :] = (
+            batch["features"][i : i + 1, :n, :] - mel_mean
+        ) * mel_std_inv
+        batch["features"][i : i + 1, n:, :] = 0
+
     audio = batch["audio"].to(device)
     features = batch["features"].to(device)
     audio_lens = batch["audio_lens"].to(device)
@@ -298,7 +312,7 @@ def prepare_input(batch: dict, tokenizer: Tokenizer, device: torch.device):
     tokens = batch["tokens"]
 
     tokens = tokenizer.tokens_to_token_ids(
-        tokens, intersperse_blank=True, add_sos=True, add_eos=True
+        tokens, intersperse_blank=True, add_sos=False, add_eos=False
     )
     tokens = k2.RaggedTensor(tokens)
     row_splits = tokens.shape.row_splits(1)
@@ -315,7 +329,7 @@ def prepare_input(batch: dict, tokenizer: Tokenizer, device: torch.device):
 
         #  features_lens[features_lens.argmax()] += pad
 
-    return audio, audio_lens, features, features_lens, tokens, tokens_lens
+    return audio, audio_lens, features, features_lens.long(), tokens, tokens_lens.long()
 
 
 def compute_validation_loss(
@@ -336,28 +350,36 @@ def compute_validation_loss(
 
     with torch.no_grad():
         for batch_idx, batch in enumerate(valid_dl):
+            if "tokens" in batch:
 
-            (
-                audio,
-                audio_lens,
-                features,
-                features_lens,
-                tokens,
-                tokens_lens,
-            ) = prepare_input(batch, tokenizer, device)
+                (
+                    audio,
+                    audio_lens,
+                    features,
+                    features_lens,
+                    tokens,
+                    tokens_lens,
+                ) = prepare_input(batch, tokenizer, device, params)
 
-            losses = get_losses(
-                {
-                    "x": tokens,
-                    "x_lengths": tokens_lens,
-                    "y": features.permute(0, 2, 1),
-                    "y_lengths": features_lens,
-                    "spks": None,  # should change it for multi-speakers
-                    "durations": None,
-                }
-            )
+                losses = get_losses(
+                    {
+                        "x": tokens,
+                        "x_lengths": tokens_lens,
+                        "y": features.permute(0, 2, 1),
+                        "y_lengths": features_lens,
+                        "spks": None,  # should change it for multi-speakers
+                        "durations": None,
+                    }
+                )
 
-            batch_size = len(batch["tokens"])
+                batch_size = len(batch["tokens"])
+            else:
+                batch_size = batch["x"].shape[0]
+                batch["x"] = batch["x"].to(device)
+                batch["x_lengths"] = batch["x_lengths"].to(device)
+                batch["y"] = batch["y"].to(device)
+                batch["y_lengths"] = batch["y_lengths"].to(device)
+                losses = get_losses(batch)
 
             loss_info = MetricsTracker()
             loss_info["samples"] = batch_size
@@ -451,24 +473,38 @@ def train_one_epoch(
         # features_lens, (N,), int32
         # tokens: List[List[str]], len(tokens) == N
 
-        batch_size = len(batch["tokens"])
-
-        audio, audio_lens, features, features_lens, tokens, tokens_lens = prepare_input(
-            batch, tokenizer, device
-        )
+        if "tokens" in batch:
+            batch_size = len(batch["tokens"])
 
+            (
+                audio,
+                audio_lens,
+                features,
+                features_lens,
+                tokens,
+                tokens_lens,
+            ) = prepare_input(batch, tokenizer, device, params)
+        else:
+            batch_size = batch["x"].shape[0]
         try:
             with autocast(enabled=params.use_fp16):
-                losses = get_losses(
-                    {
-                        "x": tokens,
-                        "x_lengths": tokens_lens,
-                        "y": features.permute(0, 2, 1),
-                        "y_lengths": features_lens,
-                        "spks": None,  # should change it for multi-speakers
-                        "durations": None,
-                    }
-                )
+                if "tokens" in batch:
+                    losses = get_losses(
+                        {
+                            "x": tokens,
+                            "x_lengths": tokens_lens,
+                            "y": features.permute(0, 2, 1),
+                            "y_lengths": features_lens,
+                            "spks": None,  # should change it for multi-speakers
+                            "durations": None,
+                        }
+                    )
+                else:
+                    batch["x"] = batch["x"].to(device)
+                    batch["x_lengths"] = batch["x_lengths"].to(device)
+                    batch["y"] = batch["y"].to(device)
+                    batch["y_lengths"] = batch["y_lengths"].to(device)
+                    losses = get_losses(batch)
 
                 loss = sum(losses.values())
 
@@ -586,6 +622,7 @@ def run(rank, world_size, args):
     params.blank_id = tokenizer.pad_id
     params.vocab_size = tokenizer.vocab_size
     params.model_args.n_vocab = params.vocab_size
+    params.model_args.n_vocab = 178
 
     logging.info(params)
     print(params)
@@ -595,7 +632,6 @@ def run(rank, world_size, args):
 
     num_param = sum([p.numel() for p in model.parameters()])
     logging.info(f"Number of parameters: {num_param}")
-    print(f"Number of parameters: {num_param}")
 
     assert params.start_epoch > 0, params.start_epoch
     checkpoints = load_checkpoint_if_available(params=params, model=model)
@@ -609,13 +645,21 @@ def run(rank, world_size, args):
     optimizer = torch.optim.Adam(model.parameters(), **params.model_args.optimizer)
 
     logging.info("About to create datamodule")
-    ljspeech = LJSpeechTtsDataModule(args)
 
-    train_cuts = ljspeech.train_cuts()
-    train_dl = ljspeech.train_dataloaders(train_cuts)
+    if False:
+        params.data_args.tokenizer = tokenizer
+        data_module = TextMelDataModule(hparams=params.data_args)
+        del params.data_args.tokenizer
+        train_dl = data_module.train_dataloader()
+        valid_dl = data_module.val_dataloader()
+    else:
+        ljspeech = LJSpeechTtsDataModule(args)
 
-    valid_cuts = ljspeech.valid_cuts()
-    valid_dl = ljspeech.valid_dataloaders(valid_cuts)
+        train_cuts = ljspeech.train_cuts()
+        train_dl = ljspeech.train_dataloaders(train_cuts)
+
+        valid_cuts = ljspeech.valid_cuts()
+        valid_dl = ljspeech.valid_dataloaders(valid_cuts)
 
     scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
     if checkpoints and "grad_scaler" in checkpoints:
@@ -625,7 +669,8 @@ def run(rank, world_size, args):
     for epoch in range(params.start_epoch, params.num_epochs + 1):
         logging.info(f"Start epoch {epoch}")
         fix_random_seed(params.seed + epoch - 1)
-        train_dl.sampler.set_epoch(epoch - 1)
+        if "sampler" in train_dl:
+            train_dl.sampler.set_epoch(epoch - 1)
 
         params.cur_epoch = epoch
 

egs/ljspeech/TTS/matcha/tts_datamodule.py

@@ -181,6 +181,8 @@ class LJSpeechTtsDataModule:
                 frame_length=1024 / sampling_rate,  # (in second),
                 frame_shift=256 / sampling_rate,  # (in second)
                 use_fft_mag=True,
+                remove_dc_offset=False,
+                preemph_coeff=0,
                 low_freq=0,
                 high_freq=8000,
                 # should be identical to n_feats in ./train.py
@@ -242,6 +244,8 @@ class LJSpeechTtsDataModule:
                 frame_length=1024 / sampling_rate,  # (in second),
                 frame_shift=256 / sampling_rate,  # (in second)
                 use_fft_mag=True,
+                remove_dc_offset=False,
+                preemph_coeff=0,
                 low_freq=0,
                 high_freq=8000,
                 # should be identical to n_feats in ./train.py
@@ -286,6 +290,8 @@ class LJSpeechTtsDataModule:
                 frame_length=1024 / sampling_rate,  # (in second),
                 frame_shift=256 / sampling_rate,  # (in second)
                 use_fft_mag=True,
+                remove_dc_offset=False,
+                preemph_coeff=0,
                 low_freq=0,
                 high_freq=8000,
                 # should be identical to n_feats in ./train.py

egs/ljspeech/TTS/prepare.sh

@@ -191,3 +191,10 @@ if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
       touch data/fbank/.ljspeech_split.done
   fi
 fi
+
+if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
+  log "Stage 9: Compute fbank mean and std (used by ./matcha)"
+  if [ ! -f ./data/fbank/cmvn.json ]; then
+    ./local/compute_fbank_statistics.py ./data/fbank/ljspeech_cuts_train.jsonl.gz ./data/fbank/cmvn.json
+  fi
+fi