push_s2t_translation.py

import json
import os
import time
from os.path import join as p_join
from tqdm import tqdm
from typing import Dict
from glob import glob

from soundfile import LibsndfileError
from datasets import Dataset, Audio, DatasetDict

# dataset config
direction_speech = os.getenv("DIRECTION_SPEECH", "enA")
direction_text = os.getenv("DIRECTION_TEXT", "jpn")
direction = f"{direction_speech}-{direction_text}"
with open(f"text.{direction}.json") as f:
    line2text = json.load(f)
cache_dir_audio = p_join("download", "audio", direction)
cache_dir_feature = p_join("download", "feature", direction)
os.makedirs(cache_dir_audio, exist_ok=True)
os.makedirs(cache_dir_feature, exist_ok=True)
line_no_start = int(os.getenv("LINE_NO_START", 0))
line_no_end = int(os.getenv("LINE_NO_END", 10000))
dataset_id = int(os.getenv("DATASET_ID", 0))
hf_org = "kotoba-tech"
hf_dataset = f"seamless-align-{direction}"


def loader(feature: str) -> Dict:
    with open(feature) as f:
        return json.load(f)


# create a dataset instance

files = {
    int(os.path.basename(i).replace(".json", "")): i for i in glob(p_join(cache_dir_feature, "*.json"))
}


def delete_audio(target_audio_file):
    if os.path.exists(target_audio_file):
        os.remove(target_audio_file)
    line_no = os.path.basename(target_audio_file).split(".")[0]
    try:
        feature_file = files[int(line_no)]
        if os.path.exists(feature_file):
            os.remove(feature_file)
    except Exception as e:
        print(e)


# remove broken audio files
features = []
audio_loader = Audio()
for i in tqdm(list(range(line_no_start, line_no_end))):
    if i in files:
        continue
    i = loader(files[i])
    i[f"{direction_text}.text"] = line2text[str(i)]
    audio_file = i[f"{direction_speech}.path"]
    start, end = i[f"{direction_speech}.duration_start"], i[f"{direction_speech}.duration_end"]
    if os.path.exists(audio_file):
        try:
            wav = audio_loader.decode_example({"path": audio_file, "bytes": None})
            if start < end < len(wav["array"]):
                features.append(i)
            else:
                delete_audio(audio_file)
        except Exception as e:
            print(e)
            delete_audio(audio_file)


print(f"features (filtered): {len(features)}")
data_dict = {f"{direction_speech}.audio": [i.pop(f"{direction_speech}.path") for i in features]}
keys = features[0].keys()
data_dict.update({k: [i[k] for i in features] for k in keys})
audio_dataset = Dataset.from_dict(data_dict)
audio_dataset = audio_dataset.cast_column(f"{direction_speech}.audio", Audio())


# trim the audio according to the duration
def clip_audio(batch):
    start = batch[f"{direction_speech}.duration_start"]
    end = batch[f"{direction_speech}.duration_end"]
    audio = batch[f"{direction_speech}.audio"]
    batch[f"{direction_speech}.audio"] = [
        {"array": a["array"][s:e], "sampling_rate": a["sampling_rate"]}
        for a, s, e in zip(audio, start, end)
    ]
    return batch


audio_dataset_valid = audio_dataset_valid.map(
    function=clip_audio,
    batched=True,
    batch_size=128,
    num_proc=1,
    desc="clipping audio based on the duration:"
)

dataset_to_push = DatasetDict({"train": audio_dataset_valid})
repo_name = f"{hf_org}/{hf_dataset}"
while True:
    try:
        dataset_to_push.push_to_hub(repo_name, config_name=f"subset_{dataset_id}")
        break
    except Exception:
        print(f"FAILED: push_to_hub on {repo_name} failed. wait 60 sec and retry soon...")
        time.sleep(60)
os.makedirs("log", exist_ok=True)
with open(f"log/pushed.line_no.{dataset_id}.json", "w") as f:
    json.dump(data_dict["line_no"], f)