Minor fixes to infer pretrained model

egs/zipvoice/README.md

@@ -21,7 +21,33 @@ ZipVoice is a high-quality zero-shot TTS model with a small model size and fast
 
 
 ## Installation
+
+* Clone icefall repository and change to zipvoice directory:
+
+```bash
+git clone https://github.com/k2-fsa/icefall.git
+cd icefall/egs/zipvoice
 ```
+
+* Create a Python virtual environment (optional but recommended):
+
+```bash
+python3 -m venv venv
+source venv/bin/activate
+```
+
+* Install the required packages:
+
+```bash
+# Install pytorch and k2.
+# If you want to use different versions, please refer to https://k2-fsa.org/get-started/k2/ for details.
+# For users in China mainland, please refer to https://k2-fsa.org/zh-CN/get-started/k2/
+
+pip install torch==2.5.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu121
+pip install k2==1.24.4.dev20250208+cuda12.1.torch2.5.1 -f https://k2-fsa.github.io/k2/cuda.html
+
+# Install other dependencies.
+pip install piper_phonemize -f https://k2-fsa.github.io/icefall/piper_phonemize.html
 pip install -r requirements.txt
 ```
 
@@ -31,12 +57,21 @@ To generate speech with our pre-trained ZipVoice or ZipVoice-Distill models, use
 
 ### 1. Inference of a single sentence:
 ```bash
+# Chinese example
 python3 zipvoice/zipvoice_infer.py \
     --model-name "zipvoice_distill" \
-    --prompt-wav prompt.wav \
-    --prompt-text "I am the transcription of the prompt wav." \
-    --text "I am the text to be synthesized." \
-    --res-wav-path result.wav
+    --prompt-wav assets/prompt-zh.wav \
+    --prompt-text "对，这就是我，万人敬仰的太乙真人，虽然有点婴儿肥，但也掩不住我逼人的帅气。" \
+    --text "欢迎使用我们的语音合成模型，希望它能给你带来惊喜！" \
+    --res-wav-path result-zh.wav
+
+# English example
+python3 zipvoice/zipvoice_infer.py \
+    --model-name "zipvoice_distill" \
+    --prompt-wav assets/prompt-en.wav \
+    --prompt-text "Some call me nature, others call me mother nature. I've been here for over four point five billion years, twenty two thousand five hundred times longer than you." \
+    --text "Welcome to use our tts model, have fun!" \
+    --res-wav-path result-en.wav
 ```
 
 ### 2. Inference of a list of sentences:
@@ -46,6 +81,7 @@ python3 zipvoice/zipvoice_infer.py \
     --test-list test.tsv \
     --res-dir results/test
 ```
+
 - `--model-name` can be `zipvoice` or `zipvoice_distill`, which are models before and after distillation, respectively.
 - Each line of `test.tsv` is in the format of `{wav_name}\t{prompt_transcription}\t{prompt_wav}\t{text}`.
 
@@ -357,4 +393,4 @@ on three test sets, i.e., LibriSpeech-PC test-clean, Seed-TTS test-en and Seed-T
       journal={arXiv preprint arXiv:2506.13053},
       year={2025},
 }
-```
+```

egs/zipvoice/scripts/prepare.sh

@@ -0,0 +1,232 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+# add icefall to PYTHONPATH
+export PYTHONPATH=../../../:$PYTHONPATH
+
+set -eou pipefail
+
+stage=0
+stop_stage=100
+
+token_type=bpe  # bpe, letter, phone
+bpe_vocab_size=500
+
+nj=32
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+
+if [ $stage -le -2 ] && [ $stop_stage -ge -2 ]; then
+
+  if [ ! -d $dl_dir/xvector_nnet_1a_libritts_clean_460 ]; then
+    log "Downloading x-vector"
+
+    git clone https://huggingface.co/datasets/zrjin/xvector_nnet_1a_libritts_clean_460 $dl_dir/xvector_nnet_1a_libritts_clean_460
+
+    mkdir -p exp/xvector_nnet_1a/
+    cp -r $dl_dir/xvector_nnet_1a_libritts_clean_460/* exp/xvector_nnet_1a/
+  fi
+
+fi
+
+if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
+  log "Stage -1: build monotonic_align lib"
+  if [ ! -d vits/monotonic_align/build ]; then
+    cd vits/monotonic_align
+    python setup.py build_ext --inplace
+    cd ../../
+  else
+    log "monotonic_align lib already built"
+  fi
+fi
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/LibriTTS,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/LibriTTS $dl_dir/LibriTTS
+  #
+  if [ ! -d $dl_dir/LibriTTS ]; then
+    lhotse download libritts $dl_dir
+  fi
+
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare LibriTTS manifest"
+  # We assume that you have downloaded the LibriTTS corpus
+  # to $dl_dir/LibriTTS
+  mkdir -p data/manifests
+  if [ ! -e data/manifests/.libritts.done ]; then
+    lhotse prepare libritts --num-jobs ${nj} $dl_dir/LibriTTS data/manifests
+    touch data/manifests/.libritts.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Compute Fbank for LibriTTS"
+  mkdir -p data/fbank
+
+  for subset in train-clean-100 train-clean-360 train-other-500 dev-clean test-clean; do
+    python local/compute_fbank.py --dataset libritts --subset ${subset}
+  done
+
+  # Here we shuffle and combine the train-clean-100, train-clean-360 and
+  # train-other-500 together to form the training set.
+  if [ ! -f data/fbank/libritts_cuts_train-all-shuf.jsonl.gz ]; then
+    cat <(gunzip -c data/fbank/libritts_cuts_train-clean-100.jsonl.gz) \
+      <(gunzip -c data/fbank/libritts_cuts_train-clean-360.jsonl.gz) \
+      <(gunzip -c data/fbank/libritts_cuts_train-other-500.jsonl.gz) | \
+      shuf | gzip -c > data/fbank/libritts_cuts_train-all-shuf.jsonl.gz
+  fi
+
+  if [ ! -f data/fbank/libritts_cuts_train-clean-460.jsonl.gz ]; then
+    cat <(gunzip -c data/fbank/libritts_cuts_train-clean-100.jsonl.gz) \
+      <(gunzip -c data/fbank/libritts_cuts_train-clean-360.jsonl.gz) | \
+      shuf | gzip -c > data/fbank/libritts_cuts_train-clean-460.jsonl.gz
+  fi
+
+  if [ ! -e data/fbank/.libritts-validated.done ]; then
+    log "Validating data/fbank for LibriTTS"
+    ./local/validate_manifest.py \
+      data/fbank/libritts_cuts_train-all-shuf.jsonl.gz
+    touch data/fbank/.libritts-validated.done
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Prepare tokens.txt"
+
+  if [ $token_type == "bpe" ] || [ $token_type == "letter" ]; then
+    if [ ! -e data/texts.txt ]; then
+      ./local/export_normalized_texts.py --output data/texts.txt \
+        --manifests data/fbank/libritts_cuts_train-all-shuf.jsonl.gz
+    fi
+  fi
+
+  if [ $token_type == "bpe" ]; then
+    mkdir -p data/lang_bpe_${bpe_vocab_size}
+    if [ ! -e data/lang_bpe_${bpe_vocab_size}/tokens.txt ]; then
+      ./local/train_bpe_model.py --transcript data/texts.txt \
+        --lang-dir data/lang_bpe_${bpe_vocab_size} \
+        --vocab-size $bpe_vocab_size
+    fi
+  fi
+
+  if [ $token_type == "phone" ]; then
+    mkdir -p data/lang_phone
+    ./local/export_tokens.py --token-type phone \
+      --output data/lang_phone/tokens.txt
+  fi
+
+  if [ $token_type == "letter" ]; then
+    mkdir -p data/lang_letter
+    ./local/export_tokens.py --token-type letter  \
+      --texts data/texts.txt \
+      --output data/lang_letter/tokens.txt
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Download and prepare librispeech-pc test clean for testing."
+
+  if [ ! -e $dl_dir/test-clean.tar.gz ]; then
+    wget https://huggingface.co/datasets/k2-fsa/LibriSpeech/resolve/main/test-clean.tar.gz -P $dl_dir
+  fi
+  # For China users.
+  if [ ! -e $dl_dir/test-clean.tar.gz ]; then
+    wget https://hf-mirror.com/datasets/k2-fsa/LibriSpeech/resolve/main/test-clean.tar.gz -P $dl_dir
+  fi
+
+  if [ ! -d $dl_dir/LibriSpeech/test-clean ]; then
+    tar -xvf $dl_dir/test-clean.tar.gz -C $dl_dir
+  fi
+
+  mkdir -p $dl_dir/LibriSpeech-PC
+  if [ ! -e $dl_dir/LibriSpeech-PC/test-clean.json ]; then
+    wget https://us.openslr.org/resources/145/manifests.tar.gz -P $dl_dir/LibriSpeech-PC
+    tar -xvf $dl_dir/LibriSpeech-PC/manifests.tar.gz -C $dl_dir/LibriSpeech-PC
+  fi
+
+  python local/compute_fbank.py --dataset librispeech --subset test-clean
+  python local/prepare_prompts_librispeech_test_clean.py
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Compute Spectrogram for LibriTTS (for VITS system)"
+  mkdir -p data/spectrogram
+  if [ ! -e data/spectrogram/.libritts.done ]; then
+    ./local/compute_spectrogram_libritts.py --sampling-rate $sampling_rate
+    touch data/spectrogram/.libritts.done
+  fi
+
+  # Here we shuffle and combine the train-clean-100, train-clean-360 and
+  # train-other-500 together to form the training set.
+  if [ ! -f data/spectrogram/libritts_cuts_train-all-shuf.jsonl.gz ]; then
+    cat <(gunzip -c data/spectrogram/libritts_cuts_train-clean-100.jsonl.gz) \
+      <(gunzip -c data/spectrogram/libritts_cuts_train-clean-360.jsonl.gz) \
+      <(gunzip -c data/spectrogram/libritts_cuts_train-other-500.jsonl.gz) | \
+      shuf | gzip -c > data/spectrogram/libritts_cuts_train-all-shuf.jsonl.gz
+  fi
+
+  # Here we shuffle and combine the train-clean-100, train-clean-360
+  # together to form the training set.
+  if [ ! -f data/spectrogram/libritts_cuts_train-clean-460.jsonl.gz ]; then
+    cat <(gunzip -c data/spectrogram/libritts_cuts_train-clean-100.jsonl.gz) \
+      <(gunzip -c data/spectrogram/libritts_cuts_train-clean-360.jsonl.gz) | \
+      shuf | gzip -c > data/spectrogram/libritts_cuts_train-clean-460.jsonl.gz
+  fi
+
+  if [ ! -e data/spectrogram/.libritts-validated.done ]; then
+    log "Validating data/spectrogram for LibriTTS"
+    ./local/validate_manifest.py \
+      data/spectrogram/libritts_cuts_train-all-shuf.jsonl.gz
+    touch data/spectrogram/.libritts-validated.done
+  fi
+fi
+
+audio_feats_dir=data/tokenized
+dataset_parts="--dataset-parts all"  # debug "-p dev-clean -p test-clean"
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Tokenize/Fbank LibriTTS for valle"
+  mkdir -p ${audio_feats_dir}
+  if [ ! -e ${audio_feats_dir}/.libritts.tokenize.done ]; then
+    python3 ./local/compute_neural_codec_and_prepare_text_tokens.py --dataset-parts "${dataset_parts}" \
+        --audio-extractor "Encodec" \
+        --batch-duration 400 \
+        --src-dir "data/manifests" \
+        --output-dir "${audio_feats_dir}"
+  fi
+  touch ${audio_feats_dir}/.libritts.tokenize.done
+
+  lhotse combine \
+    ${audio_feats_dir}/libritts_cuts_train-clean-100.jsonl.gz \
+    ${audio_feats_dir}/libritts_cuts_train-clean-360.jsonl.gz \
+    ${audio_feats_dir}/libritts_cuts_train-other-500.jsonl.gz \
+    ${audio_feats_dir}/cuts_train.jsonl.gz
+  lhotse copy \
+    ${audio_feats_dir}/libritts_cuts_dev-clean.jsonl.gz \
+    ${audio_feats_dir}/cuts_dev.jsonl.gz
+  lhotse copy \
+    ${audio_feats_dir}/libritts_cuts_test-clean.jsonl.gz \
+    ${audio_feats_dir}/cuts_test.jsonl.gz
+fi

egs/zipvoice/scripts/prepare_libritts.sh

@@ -85,4 +85,4 @@ if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
   if [ ! -e data/tokens_libritts.txt ]; then
     ./local/prepare_token_file_libritts.py --tokens data/tokens_libritts.txt
   fi
-fi
+fi

egs/zipvoice/scripts/run_eval.sh

@@ -0,0 +1,75 @@
+#!/usr/bin/env bash
+
+export PYTHONPATH=../../../:$PYTHONPATH
+
+stage=1
+stop_stage=10
+generated_wav_path="flow-matching/exp/generated_wavs"
+
+. shared/parse_options.sh || exit 1
+
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+if [ $stage -le -2 ] && [ $stop_stage -ge -2 ]; then
+  log "Stage -2: Install dependencies and download models"
+
+  pip install -r requirements-eval.txt
+  pip install git+https://github.com/sarulab-speech/UTMOSv2.git
+  modelscope download --model k2-fsa/TTS_eval_models --local_dir TTS_eval_models
+fi
+
+
+if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
+  log "Stage -1: Prepare evaluation data."
+
+  mkdir -p data/reference/librispeech-test-clean
+
+  gunzip -c data/fbank/librispeech_cuts_with_prompts_test-clean.jsonl.gz | \
+    jq -r '"\(.recording.sources[0].source)"' | \
+    awk '{split($1, a, "/"); cmd="cp "$1" data/reference/librispeech-test-clean/"a[length(a)]; print cmd; system(cmd)}'
+
+
+  mkdir -p data/reference/librispeech-test-clean-prompt
+  gunzip -c data/fbank/librispeech_cuts_with_prompts_test-clean.jsonl.gz | \
+    jq -r '"\(.custom.prompt.recording.sources[0].source) \(.recording.sources[0].source)"' | \
+    awk '{split($2, a, "/"); cmd="cp "$1" data/reference/librispeech-test-clean-prompt/"a[length(a)]; print cmd; system(cmd)}'
+fi
+
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Evaluate the model with FSD."
+
+  python local/evaluate_fsd.py --real-path data/reference/librispeech-test-clean \
+    --eval-path $generated_wav_path
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Evaluate the model with SIM."
+
+  python local/evaluate_sim.py --real-path data/reference/librispeech-test-clean \
+    --eval-path $generated_wav_path
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Evaluate the model with UTMOS."
+
+  python local/evaluate_utmos.py --wav-path $generated_wav_path
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Evaluate the model with UTMOSv2."
+
+  python local/evaluate_utmosv2.py --wav-path $generated_wav_path
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Evaluate the model with WER."
+
+  python local/evaluate_wer_hubert.py --wav-path $generated_wav_path \
+    --decode-path $generated_wav_path/decode 
+fi

egs/zipvoice/zipvoice/zipvoice_infer.py

@@ -23,7 +23,7 @@ This script generates speech with our pre-trained ZipVoice or
 Usage:
 
 Note: If you having trouble connecting to HuggingFace,
-    you try switch endpoint to mirror site:
+    try switching endpoint to mirror site:
 
 export HF_ENDPOINT=https://hf-mirror.com
 
@@ -55,7 +55,6 @@ import os
 
 import numpy as np
 import safetensors.torch
-import soundfile as sf
 import torch
 import torch.nn as nn
 import torchaudio
@@ -115,15 +114,20 @@ def get_parser():
         "--res-dir",
         type=str,
         default="results",
-        help="Path name of the generated wavs dir, "
-        "used when decdode-list is not None",
+        help="""
+        Path name of the generated wavs dir,
+        used when test-list is not None
+        """,
     )
 
     parser.add_argument(
         "--res-wav-path",
         type=str,
         default="result.wav",
-        help="Path name of the generated wav path, " "used when decdode-list is None",
+        help="""
+        Path name of the generated wav path,
+        used when test-list is None
+        """,
     )
 
     parser.add_argument(
@@ -456,8 +460,7 @@ def generate_sentence(
     # Adjust wav volume if necessary
     if prompt_rms < target_rms:
         wav = wav * prompt_rms / target_rms
-    wav = wav[0].cpu().numpy()
-    sf.write(save_path, wav, sampling_rate)
+    torchaudio.save(save_path, wav.cpu(), sample_rate=sampling_rate)
 
     return metrics