faster-build.py

import argparse
import csv
from datetime import datetime
import os
import re
import shutil
import tarfile
import tempfile
from tqdm import tqdm
import torchaudio
from pydub import AudioSegment
import requests
from pocketbase import PocketBase
from torchaudio.transforms import Resample
from concurrent.futures import ThreadPoolExecutor

parser = argparse.ArgumentParser(description="Command description.")

pb = PocketBase('https://pocketbase.nenadb.dev/')

def contains_interruption(transcription: str, translation: str) -> bool:
    boundaries = r"[\s\-꞊ˈ…,\.?!]|$"
    languages = r"(A|Az|E|H|K|P|R)"

    # Check if transcription is just a string enclosed with parenthesis
    if re.fullmatch(r'\(.*\)', transcription):
        return True

    # Check if transcription contains any language abbreviation followed by a boundary
    pattern = f'{languages}(?={boundaries})'
    if re.search(pattern, transcription):
        return True

    # Check if translation contains square brackets
    if '[' in translation and ']' in translation:
        return True

    # If none of the above conditions are met, return False
    return False

def process_example(example, dialect, split, audio_dir_path, transcripts, stats):
    audio_file_name = f"nena_speech_{example.id}.mp3"
    audio_file_path = os.path.join(audio_dir_path, audio_file_name)

    if not os.path.exists(audio_file_path):
        audio_url = pb.get_file_url(example, example.speech, {})
        response = requests.get(audio_url)
        with tempfile.NamedTemporaryFile() as f:
            f.write(response.content)
            f.flush()
            waveform, sample_rate = torchaudio.load(f.name)
        resampler = Resample(orig_freq=sample_rate, new_freq=48000)
        resampled_waveform = resampler(waveform)
        torchaudio.save(audio_file_path, resampled_waveform, 48000, format="mp3")

    transcripts.append({
        'transcription': example.transcription,
        'translation': example.translation,
        'locale': example.locale,
        'proficiency': example.proficiency,
        'age': example.age,
        'crowdsourced': example.crowdsourced,
        'interrupted': contains_interruption(example.transcription, example.translation),
        'client_id': example.speaker,
        'path': audio_file_name,
    })

    dialect_stats = stats["dialects"][dialect]

    stats["totalExamples"] += 1
    dialect_stats["totalExamples"] += 1
    if example.translation:
        stats["examplesTranslated"] += 1
        dialect_stats["examplesTranslated"] += 1
    if example.transcription:
        stats["durationLabeled"] += 0
        dialect_stats["durationLabeled"] += 0
    else:
        stats["durationUnlabeled"] += 0
        dialect_stats["durationUnlabeled"] += 0

    dialect_stats["buckets"][split] += 1
    dialect_stats["speakers"].add(example.speaker)
    dialect_stats["splits"]["proficiency"][example.proficiency] = dialect_stats["splits"]["proficiency"].get(example.proficiency, 0) + 1
    dialect_stats["splits"]["age"][example.age] = dialect_stats["splits"]["age"].get(example.age, 0) + 1
    dialect_stats["splits"]["locale"][example.locale] = dialect_stats["splits"]["locale"].get(example.locale, 0) + 1
    if example.crowdsourced:
        dialect_stats["splits"]["crowdsourced"] += 1

def build_dataset(version, test_split=0.10, dev_split=0.10):
    dialects = pb.collection("dialects").get_full_list(query_params={
        "sort": "name",
    })

    dialects = {
        dialect.name.lower(): dialect.name
        for dialect in dialects
    }

    examples = pb.collection("examples").get_full_list(query_params={
        "expand": "dialect",
        "filter": "validated=true",
    })

    stats = {
        "dialects": {
            dialect : {
                "buckets": {
                    "dev": 0,
                    "test": 0,
                    "train": 0,
                },
                "splits": {
                    "proficiency": {},
                    "age": {},
                    "locale": {},
                    "crowdsourced": 0,
                },
                "speakers": set(),
                "totalExamples": 0,
                "examplesTranslated": 0,
                "durationLabeled": 0,
                "durationUnlabeled": 0,
            }
            for dialect in dialects.keys()
        },
        "totalExamples": 0,
        "examplesTranslated": 0,
        "durationLabeled": 0,
        "durationUnlabeled": 0,
        "version": version,
        "date": datetime.now().strftime("%Y-%m-%d"),
        "name": "NENA Speech Dataset",
        "multilingual": True,
    }

    def split_examples(examples):
        test_end = int(test_split * len(examples))
        dev_end = int((dev_split + test_split) * len(examples))
        
        return {
            'test': examples[:test_end],
            'dev': examples[test_end:dev_end],
            'train': examples[dev_end:],
        }

    subsets = {
        dialect: split_examples([
            example for example in examples
            if example.expand['dialect'].name.lower() == dialect
        ])
        for dialect in dialects.keys()
    }

    with tqdm(total=len(examples)) as pbar:
        for dialect, subset in subsets.items():
            for split, examples in subset.items():
                audio_dir_path = os.path.join("audio", dialect, split)
                audio_tar_path = f"{audio_dir_path}.tar"
                if os.path.exists(audio_tar_path):
                    with tarfile.open(audio_tar_path, "r") as tar:
                        tar.extractall(path=os.path.join("audio", dialect))
                else:
                    os.makedirs(audio_dir_path, exist_ok=True)
                
                transcripts = []
                transcript_dir_path = os.path.join("transcript", dialect)
                os.makedirs(transcript_dir_path, exist_ok=True)
                
                # Parallelize processing examples
                with ThreadPoolExecutor() as executor:
                    futures = [
                        executor.submit(process_example, example, dialect, split, audio_dir_path, transcripts, stats)
                        for example in examples
                    ]
                    for future in futures:
                        pbar.update(1)
                        future.result()

                pbar.set_description(f"Saving audios ({dialect}/{split})")
                audio_tar_path = f"{audio_dir_path}.tar"
                with tarfile.open(audio_tar_path, 'w') as tar:
                    tar.add(audio_dir_path, arcname=os.path.basename(audio_dir_path))

                pbar.set_description(f"Saving transcripts ({dialect}/{split})")

                with open(os.path.join(transcript_dir_path, f"{split}.tsv"), 'w', newline='') as f:
                    fieldnames = [] if len(transcripts) == 0 else transcripts[0].keys()
                    writer = csv.DictWriter(f, fieldnames=fieldnames, delimiter='\t')
                    writer.writeheader()
                    writer.writerows(transcripts)

                shutil.rmtree(audio_dir_path) 
            stats["dialects"][dialect]["speakers"] = len(stats["dialects"][dialect]["speakers"])

    with open('dialects.py', 'w') as f:
        python_code = f'DIALECTS = {repr(dialects)}\n'
        f.write(python_code)

    with open('release_stats.py', 'w') as f:
        python_code = f'STATS = {repr(stats)}\n'
        f.write(python_code)


if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Generate text from prompt")

    parser.add_argument(
        "-b",
        "--build",
        action="store_true",
        help="Download text prompts from GCS bucket",
    )

    parser.add_argument(
        "-v",
        "--version",
        type=str,
        default="1.0.0",
        help="Download text prompts from GCS bucket",
    )

    args = parser.parse_args()

    if args.build:
        build_dataset(version=args.version)