code-switch_chunks / data_preperation.py
1uckyan's picture
Upload data_preperation.py
1ba8d1f verified
raw
history blame
7.29 kB
import os
import re
import json
import shutil
import argparse
import librosa
import soundfile as sf
from tqdm import tqdm
# ARGS CONFIGURATION
def parse_args():
parser = argparse.ArgumentParser(description="Reproduce mixed Code-Switching Dataset.")
parser.add_argument("--secomicsc_root", type=str, required=True,
help="Path to 'ASR-SECoMiCSC' folder (must contain TXT and WAV subfolders).")
parser.add_argument("--dev_root", type=str, required=True,
help="Path to 'ASR-DevCECoMiCSC' folder (must contain TXT and WAV subfolders).")
parser.add_argument("--cs_dialogue_root", type=str, required=True,
help="Path to CS-Dialogue 'short_wav' folder (must contain SCRIPT and WAVE).")
parser.add_argument("--output_dir", type=str, default="./CS_chunks_Dataset",
help="Directory to save processed audio and metadata.")
return parser.parse_args()
# CONSTANTS
TARGET_SR = 16000
MIN_DURATION = 5.0
MAX_DURATION = 15.0
MAX_GAP = 1.8
NOISE_TAGS = ["[ENS]", "[NPS]", "[SONANT]", "[*]", "[LAUGHTER]"]
# LEGACY PROCESSING LOGIC
def parse_legacy_line(line):
line = line.strip()
if not line: return None
m = re.match(r"\[([\d.]+),([\d.]+)\]\s+(.*)", line)
if not m: return None
start, end = float(m.group(1)), float(m.group(2))
rest = m.group(3).split()
if len(rest) < 2: return None
text = " ".join(rest[2:]) if len(rest) >= 3 else rest[-1]
is_noise = any(tag in text for tag in NOISE_TAGS)
return {"start": start, "end": end, "text": text, "is_noise": is_noise}
def process_legacy(dataset_name, specific_root_path, meta_f, audio_out_root):
print(f"Processing Legacy: {dataset_name}...")
txt_dir = os.path.join(specific_root_path, "TXT")
wav_dir = os.path.join(specific_root_path, "WAV")
# audio/SECoMiCSC
sub_dir = os.path.join(audio_out_root, dataset_name)
os.makedirs(sub_dir, exist_ok=True)
if not os.path.exists(txt_dir):
print(f"Skipping {dataset_name}: 'TXT' folder not found inside {specific_root_path}")
return
files = [f for f in os.listdir(txt_dir) if f.endswith(".txt")]
for txt_file in tqdm(files, desc=dataset_name):
wav_file = txt_file.replace(".txt", ".wav")
wav_path = os.path.join(wav_dir, wav_file)
txt_path = os.path.join(txt_dir, txt_file)
if not os.path.exists(wav_path): continue
try:
audio, sr = librosa.load(wav_path, sr=TARGET_SR, mono=True)
except: continue
with open(txt_path, encoding="utf-8") as f:
segments = [parse_legacy_line(l) for l in f if parse_legacy_line(l)]
segments.sort(key=lambda x: x["start"])
buffer = []
buffer_start = None
last_end = None
def flush():
nonlocal buffer, buffer_start
if not buffer: return
start_t = buffer_start
end_t = buffer[-1]["end"]
if int(start_t * sr) >= len(audio) or int(end_t * sr) > len(audio): return
chunk = audio[int(start_t * sr): int(end_t * sr)]
dur = len(chunk) / sr
if dur < 0.5 or dur > MAX_DURATION: return
texts = [s["text"] for s in buffer if not s["is_noise"]]
if not texts: return
# Save Chunk
fname = f"{dataset_name}_{os.path.basename(wav_path)[:-4]}_{int(start_t*100)}_{int(end_t*100)}.wav"
out_path = os.path.join(sub_dir, fname)
sf.write(out_path, chunk, sr)
# Write Metadata
meta_f.write(json.dumps({
"file_name": f"audio/{dataset_name}/{fname}",
"sentence": " ".join(texts),
"duration": round(dur, 2),
"source": dataset_name
}, ensure_ascii=False) + "\n")
for seg in segments:
if not buffer:
if seg["is_noise"]: continue
buffer, buffer_start = [seg], seg["start"]
last_end = seg["end"]
continue
gap = seg["start"] - last_end
est_dur = seg["end"] - buffer_start
if gap > MAX_GAP or est_dur > MAX_DURATION:
flush()
buffer = [] if seg["is_noise"] else [seg]
buffer_start = seg["start"] if buffer else None
else:
buffer.append(seg)
last_end = seg["end"]
flush()
# CS-DIALOGUE PROCESSING LOGIC
def process_cs_dialogue(source_root, meta_f, audio_out_root):
DATASET_NAME = "CS_Dialogue"
script_dir = os.path.join(source_root, "SCRIPT")
wave_root = os.path.join(source_root, "WAVE", "C0")
sub_dir = os.path.join(audio_out_root, DATASET_NAME)
os.makedirs(sub_dir, exist_ok=True)
if not os.path.exists(script_dir):
print(f"CS-Dialogue SCRIPT dir not found: {script_dir}")
return
txt_files = [f for f in os.listdir(script_dir) if f.endswith(".txt")]
for txt_file in tqdm(txt_files, desc=DATASET_NAME):
txt_path = os.path.join(script_dir, txt_file)
session_id = os.path.splitext(txt_file)[0]
src_audio_folder = os.path.join(wave_root, session_id)
if not os.path.exists(src_audio_folder): continue
with open(txt_path, 'r', encoding='utf-8') as f:
for line in f:
line = line.strip()
if not line: continue
parts = line.split(maxsplit=2)
if len(parts) < 3: continue
fname_raw, tag, text = parts[0], parts[1], parts[2]
if tag != "<MIX>": continue
if not fname_raw.endswith(".wav"): fname_raw += ".wav"
src_wav = os.path.join(src_audio_folder, fname_raw)
if os.path.exists(src_wav):
dst_wav = os.path.join(sub_dir, fname_raw)
shutil.copy2(src_wav, dst_wav)
try:
dur = librosa.get_duration(path=dst_wav)
except:
dur = 0.0
meta_f.write(json.dumps({
"file_name": f"audio/{DATASET_NAME}/{fname_raw}",
"sentence": text,
"duration": round(dur, 2),
"source": DATASET_NAME,
"original_tag": tag
}, ensure_ascii=False) + "\n")
# MAIN ENTRY
if __name__ == "__main__":
args = parse_args()
audio_out = os.path.join(args.output_dir, "audio")
meta_path = os.path.join(args.output_dir, "metadata.jsonl")
os.makedirs(audio_out, exist_ok=True)
with open(meta_path, 'w', encoding='utf-8') as mf:
process_legacy("SECoMiCSC", args.secomicsc_root, mf, audio_out)
process_legacy("DevCECoMiCSC", args.dev_root, mf, audio_out)
process_cs_dialogue(args.cs_dialogue_root, mf, audio_out)
print(f"\nAll Done! Dataset ready at: {args.output_dir}")