How to train

[Shazzum]

I am trying to train based on librispeech. I am unsure the exact way to do this. The instructions on how to prepare the files are kinda confusing to me. Can anyone explain how I can do this? I am new to coding. Thanks!

[SeeWhysimon]

1. Create a new folder data under root, download and unzip LJSpeech-1.1 to data.
2. Create a script named prepare.py under data, and the content should be

# prepare.py
import os
import random
import shutil

# === 配置 ===
metadata_path = "LJSpeech-1.1/metadata.csv"
wavs_dir = "LJSpeech-1.1/wavs"
output_train = "LJSpeech-1.1/train.txt"
output_val = "LJSpeech-1.1/val.txt"

# === 读取 metadata.csv 文件 ===
with open(metadata_path, "r", encoding="utf-8") as f:
    lines = f.readlines()

# === 打乱数据 ===
random.shuffle(lines)

# === 拆分 train / val（5% 作为 val）===
total = len(lines)
val_count = total // 20
val_lines = lines[:val_count]
train_lines = lines[val_count:]

# === 格式转换为 absolute_path|text 格式 ===
def format_lines(data_lines):
    formatted = []
    for line in data_lines:
        parts = line.strip().split("|")
        if len(parts) >= 3:
            wav_name, _, text = parts
            # ✅ 使用绝对路径
            wav_path = os.path.abspath(os.path.join(wavs_dir, f"{wav_name}.wav")).replace("\\", "/")
            formatted.append(f"{wav_path}|{text}")
    return formatted

train_formatted = format_lines(train_lines)
val_formatted = format_lines(val_lines)

# === 写入文件 ===
with open(output_train, "w", encoding="utf-8") as f:
    f.write("\n".join(train_formatted))

with open(output_val, "w", encoding="utf-8") as f:
    f.write("\n".join(val_formatted))

# === 提示完成 ===
print(f"✅ Successfully created {output_train} and {output_val}")
print(f"🔢 Total: {total}, Train: {len(train_formatted)}, Val: {len(val_formatted)}")

3. Run following commands

cd data
python prepare.py

would create necessary files like train.txt for training.

4. Change training parameters in configs/train.yaml, configs/trainer/default.yaml and etc. (Therea are too many params need to modify and you can ask gpt by copying terminal outputs, it's very easy). Gradually you would run the training process properly.