Datasets:
Upload 4 files
Browse filesuploaded some new files:
- clean_up_csv.py to clean up the original allignment csv for easier use
- libris2s_dataset.py: a pytorch dataset class as a starting point for easy loading
- a notebook to test the dataset class
- clean_up_csv.py +42 -0
- data_example.ipynb +0 -0
- libris2s_dataset.py +95 -0
- requirements.txt +8 -5
clean_up_csv.py
ADDED
|
@@ -0,0 +1,42 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import pandas as pd
|
| 2 |
+
import os
|
| 3 |
+
|
| 4 |
+
# Read the original CSV
|
| 5 |
+
alignment = pd.read_csv("alignments/all_de_en_alligned.csv", index_col=0)
|
| 6 |
+
|
| 7 |
+
# Create mapping of book numbers to English folder names
|
| 8 |
+
en_folder_map = {}
|
| 9 |
+
for folder in os.listdir("EN"):
|
| 10 |
+
book_id = folder.split('.')[0]
|
| 11 |
+
en_folder_map[book_id] = folder
|
| 12 |
+
|
| 13 |
+
# Function to construct full German audio path
|
| 14 |
+
def get_de_path(row):
|
| 15 |
+
if "67" in row['book']:
|
| 16 |
+
return os.path.join("DE","67.frankenstein_de_1211_librivox_newly_alligned", "sentence_level_audio", row['DE_audio'])
|
| 17 |
+
return os.path.join("DE", row['book'], "sentence_level_audio", row['DE_audio'])
|
| 18 |
+
|
| 19 |
+
# Function to construct full English audio path
|
| 20 |
+
def get_en_path(row):
|
| 21 |
+
book_id = str(row['book_id'])
|
| 22 |
+
if book_id in en_folder_map:
|
| 23 |
+
return os.path.join("EN", en_folder_map[book_id], "sentence_level_audio", row['EN_audio'] + ".wav")
|
| 24 |
+
return None
|
| 25 |
+
|
| 26 |
+
# Update paths in the DataFrame
|
| 27 |
+
alignment['DE_audio'] = alignment.apply(get_de_path, axis=1)
|
| 28 |
+
alignment['EN_audio'] = alignment.apply(get_en_path, axis=1)
|
| 29 |
+
|
| 30 |
+
# Drop the 'book' column since paths are now complete
|
| 31 |
+
alignment = alignment.drop('book', axis=1)
|
| 32 |
+
|
| 33 |
+
# Drop rows where EN_audio path couldn't be constructed (book_id not found)
|
| 34 |
+
alignment = alignment.dropna(subset=['EN_audio'])
|
| 35 |
+
|
| 36 |
+
# Save the cleaned up csv
|
| 37 |
+
alignment.to_csv("alignments/all_de_en_alligned_cleaned.csv", index=False)
|
| 38 |
+
|
| 39 |
+
print(f"Saved cleaned CSV with {len(alignment)} rows")
|
| 40 |
+
print("\nFirst few rows of cleaned CSV:")
|
| 41 |
+
print(alignment.head())
|
| 42 |
+
|
data_example.ipynb
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
libris2s_dataset.py
ADDED
|
@@ -0,0 +1,95 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
import torch
|
| 3 |
+
import pandas as pd
|
| 4 |
+
import torchaudio
|
| 5 |
+
from torch.utils.data import Dataset
|
| 6 |
+
from typing import List, Optional
|
| 7 |
+
|
| 8 |
+
class Libris2sDataset(torch.utils.data.Dataset):
|
| 9 |
+
def __init__(self, data_dir: str, split: str, transform=None, book_ids: Optional[List[str]]=None):
|
| 10 |
+
"""
|
| 11 |
+
Initialize the LibriS2S dataset.
|
| 12 |
+
|
| 13 |
+
Args:
|
| 14 |
+
data_dir (str): Root directory containing the dataset
|
| 15 |
+
split (str): Path to the CSV file containing alignments
|
| 16 |
+
transform (callable, optional): Optional transform to be applied on the audio
|
| 17 |
+
book_ids (List[str], optional): List of book IDs to include. If None, includes all books.
|
| 18 |
+
Example: ['9', '10', '11'] will only load these books.
|
| 19 |
+
"""
|
| 20 |
+
self.data_dir = data_dir
|
| 21 |
+
self.transform = transform
|
| 22 |
+
self.book_ids = set(book_ids) if book_ids is not None else None
|
| 23 |
+
|
| 24 |
+
# Load alignment CSV file
|
| 25 |
+
self.alignments = pd.read_csv(split)
|
| 26 |
+
|
| 27 |
+
# Create lists to store paths and metadata
|
| 28 |
+
self.de_audio_paths = []
|
| 29 |
+
self.en_audio_paths = []
|
| 30 |
+
self.de_transcripts = []
|
| 31 |
+
self.en_transcripts = []
|
| 32 |
+
self.alignment_scores = []
|
| 33 |
+
|
| 34 |
+
# Process each entry in the alignments
|
| 35 |
+
for _, row in self.alignments.iterrows():
|
| 36 |
+
# Get book ID from the path
|
| 37 |
+
book_id = str(row['book_id'])
|
| 38 |
+
|
| 39 |
+
# Skip if book_id is not in the filtered set
|
| 40 |
+
if self.book_ids is not None and book_id not in self.book_ids:
|
| 41 |
+
continue
|
| 42 |
+
|
| 43 |
+
# Get full paths from CSV
|
| 44 |
+
de_audio = os.path.join(data_dir, row['DE_audio'])
|
| 45 |
+
en_audio = os.path.join(data_dir, row['EN_audio'])
|
| 46 |
+
|
| 47 |
+
# Only add if both audio files exist
|
| 48 |
+
if os.path.exists(de_audio) and os.path.exists(en_audio):
|
| 49 |
+
self.de_audio_paths.append(de_audio)
|
| 50 |
+
self.en_audio_paths.append(en_audio)
|
| 51 |
+
self.de_transcripts.append(row['DE_transcript'])
|
| 52 |
+
self.en_transcripts.append(row['EN_transcript'])
|
| 53 |
+
self.alignment_scores.append(float(row['score']))
|
| 54 |
+
else:
|
| 55 |
+
print(f"Skipping {de_audio} or {en_audio} because they don't exist")
|
| 56 |
+
|
| 57 |
+
def __len__(self):
|
| 58 |
+
"""Return the number of items in the dataset."""
|
| 59 |
+
return len(self.de_audio_paths)
|
| 60 |
+
|
| 61 |
+
def __getitem__(self, idx):
|
| 62 |
+
"""
|
| 63 |
+
Get a single item from the dataset.
|
| 64 |
+
|
| 65 |
+
Args:
|
| 66 |
+
idx (int): Index of the item to get
|
| 67 |
+
|
| 68 |
+
Returns:
|
| 69 |
+
dict: A dictionary containing:
|
| 70 |
+
- de_audio: German audio waveform
|
| 71 |
+
- de_sample_rate: German audio sample rate
|
| 72 |
+
- en_audio: English audio waveform
|
| 73 |
+
- en_sample_rate: English audio sample rate
|
| 74 |
+
- de_transcript: German transcript
|
| 75 |
+
- en_transcript: English transcript
|
| 76 |
+
- alignment_score: Alignment score between the pair
|
| 77 |
+
"""
|
| 78 |
+
# Load audio files
|
| 79 |
+
de_audio, de_sr = torchaudio.load(self.de_audio_paths[idx])
|
| 80 |
+
en_audio, en_sr = torchaudio.load(self.en_audio_paths[idx])
|
| 81 |
+
|
| 82 |
+
# Apply transforms if specified
|
| 83 |
+
if self.transform:
|
| 84 |
+
de_audio = self.transform(de_audio)
|
| 85 |
+
en_audio = self.transform(en_audio)
|
| 86 |
+
|
| 87 |
+
return {
|
| 88 |
+
'de_audio': de_audio,
|
| 89 |
+
'de_sample_rate': de_sr,
|
| 90 |
+
'en_audio': en_audio,
|
| 91 |
+
'en_sample_rate': en_sr,
|
| 92 |
+
'de_transcript': self.de_transcripts[idx],
|
| 93 |
+
'en_transcript': self.en_transcripts[idx],
|
| 94 |
+
'alignment_score': self.alignment_scores[idx]
|
| 95 |
+
}
|
requirements.txt
CHANGED
|
@@ -1,5 +1,8 @@
|
|
| 1 |
-
aeneas=1.7.3.0
|
| 2 |
-
pandas>=1.1.4
|
| 3 |
-
pydub=0.24.1
|
| 4 |
-
beautifulsoup4=4.9.3
|
| 5 |
-
requests=2.25.1
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
aeneas=1.7.3.0
|
| 2 |
+
pandas>=1.1.4
|
| 3 |
+
pydub=0.24.1
|
| 4 |
+
beautifulsoup4=4.9.3
|
| 5 |
+
requests=2.25.1
|
| 6 |
+
torch>=2.0.0
|
| 7 |
+
torchaudio>=2.0.0
|
| 8 |
+
soundfile
|