Update README.md

import os
import soundfile as sf
import pandas as pd
import librosa
import io # Import the io module
from datasets import load_dataset

# --- Configuration (Keep these the same) ---
DATASET_NAME = "Bgeorge/CNA_AudioDataset"
TARGET_SAMPLING_RATE = 22050
BASE_OUTPUT_DIR = "data"
WAVES_DIR = os.path.join(BASE_OUTPUT_DIR, "waves")
os.makedirs(WAVES_DIR, exist_ok=True)

# --- Load Dataset ---
print("Loading dataset...")
ds = load_dataset(DATASET_NAME)


# --- Revised Processing Function ---
def process_split(split_name, dataset_split):
"""Saves audio files and generates metadata for a given dataset split."""
print(f"\nProcessing {split_name} split...")
metadata = []

for i, example in enumerate(dataset_split):

# 1. MANUAL DECODING: Read the raw audio bytes
# The 'audio' column contains a dictionary with the 'bytes' key.
raw_audio_bytes = example['audio']['bytes']

# Use io.BytesIO to treat the raw bytes as a file object
audio_stream = io.BytesIO(raw_audio_bytes)

# Use soundfile.read to decode the audio stream into an array and SR
# This completely bypasses the datasets internal decoder (torchcodec)
try:
audio_array, original_sr = sf.read(audio_stream)
except Exception as e:
print(f"Skipping example {i} due to decoding error: {e}")
continue # Skip to the next example on error

text = example['text']
speaker_id = example.get('speaker_name', 'default_speaker')

# 2. Resample the audio using Librosa
if original_sr != TARGET_SAMPLING_RATE:
# Resampling requires the original sampling rate for librosa
audio_array = librosa.resample(
audio_array,
orig_sr=original_sr,
target_sr=TARGET_SAMPLING_RATE
)

# 3. Define output file path and name
file_name = f"{speaker_id}_{i:06d}.wav"
full_path = os.path.join(WAVES_DIR, file_name)
relative_path_for_csv = os.path.join("waves", file_name)

# 4. Save audio file
sf.write(full_path, audio_array, TARGET_SAMPLING_RATE, format='WAV')

# 5. Collect metadata
metadata.append({
'file_name': relative_path_for_csv,
'text': text,
'speaker_id': speaker_id
})

if (i + 1) % 1000 == 0:
print(f" Processed {i+1} examples...")

# 6. Create and save the metadata CSV
df_metadata = pd.DataFrame(metadata)
csv_filename = f"metadata{split_name}.csv"
csv_path = os.path.join(BASE_OUTPUT_DIR, csv_filename)
df_metadata.to_csv(csv_path, sep='|', index=False, header=False)

print(f"Finished processing {len(df_metadata)} examples for {split_name}.")
print(f"Metadata saved to: {csv_path}")


# --- Run Processing for Splits (Same as before) ---
if 'train' in ds:
process_split('train', ds['train'])

eval_split = ds.get('test') or ds.get('validation')
if eval_split is not None:
process_split('eval', eval_split)
else:
print("Warning: Neither 'test' nor 'validation' split found for evaluation.")

print("\n✅ Dataset conversion complete.")

README.md

@@ -1,29 +1,31 @@
----
-dataset_info:
-  features:
-  - name: audio
-    dtype:
-      audio:
-        sampling_rate: 16000
-        decode: false
-  - name: text
-    dtype: string
-  - name: speaker_name
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 6778898640
-    num_examples: 30504
-  - name: test
-    num_bytes: 750280661
-    num_examples: 3390
-  download_size: 7125544218
-  dataset_size: 7529179301
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
-  - split: test
-    path: data/test-*
----
+---
+dataset_info:
+  features:
+  - name: audio
+    dtype:
+      audio:
+        sampling_rate: 16000
+        decode: false
+  - name: text
+    dtype: string
+  - name: speaker_name
+    dtype: string
+  splits:
+  - name: train
+    num_bytes: 6778898640
+    num_examples: 30504
+  - name: test
+    num_bytes: 750280661
+    num_examples: 3390
+  download_size: 7125544218
+  dataset_size: 7529179301
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: data/train-*
+  - split: test
+    path: data/test-*
+language:
+- ro
+---