BEST-RQ-2 / audio-embeddings /src /data /audio_utils.py
ltuncay's picture
Submission to the Interspeech 2026 Audio Encoder Capability Challenge
eca55dc verified
Raw
History Blame Contribute Delete
7.08 kB
from typing import Any, Dict, List, Optional, Sequence
import numpy as np
import torch
import torchaudio
class DatasetResamplerCropper:
"""
Resample and optionally crop a waveform.
Maintains a cache of resamplers for different source sampling rates to optimize instantiation.
Args:
target_sr (int): Target sampling rate.
max_length (Optional[int]): Maximum length in samples (at target_sr).
"""
def __init__(self, target_sr: int, max_length: Optional[int] = None):
self.target_sr = target_sr
self.max_length = max_length
self.resamplers: Dict[int, torchaudio.transforms.Resample] = {}
def forward(self, waveform: torch.Tensor, source_sr: int) -> torch.Tensor:
"""
Args:
waveform (torch.Tensor): Tensor of shape [T] or [C, T].
source_sr (int): Source sampling rate.
Returns:
torch.Tensor: Processed waveform tensor.
"""
# Resampling and Cropping Logic
if source_sr != self.target_sr:
# We need to resample.
# Optimization: Crop in source domain first if we have a max_length
if self.max_length is not None:
# Calculate required source length to get max_length in target domain
# Add a small buffer to avoid rounding issues
crop_len_source = round(self.max_length * source_sr / self.target_sr)
if waveform.shape[-1] > crop_len_source:
max_start = waveform.shape[-1] - crop_len_source
start = np.random.randint(0, max_start + 1)
waveform = waveform[..., start : start + crop_len_source]
# Resample
if source_sr not in self.resamplers:
self.resamplers[source_sr] = torchaudio.transforms.Resample(
source_sr, self.target_sr
)
resampler = self.resamplers[source_sr]
waveform = resampler(waveform)
# Now handle max_length (trim if we cropped with buffer, or if it was already long enough)
if self.max_length is not None and waveform.shape[-1] > self.max_length:
waveform = waveform[..., : self.max_length]
else:
# No resampling, just standard random crop
if self.max_length is not None and waveform.shape[-1] > self.max_length:
max_start = waveform.shape[-1] - self.max_length
start = np.random.randint(0, max_start + 1)
waveform = waveform[..., start : start + self.max_length]
return waveform
def __call__(self, waveform: torch.Tensor, source_sr: int) -> torch.Tensor:
return self.forward(waveform, source_sr)
def collate_audio_batch(
batch: List[Dict[str, Any]],
waveform_key: str = "waveform",
mode: str = "pad", # "pad" or "truncate"
stack_waveforms: bool = True,
pad_value: float = 0.0,
include_keys: Optional[Sequence[str]] = None,
exclude_keys: Optional[Sequence[str]] = None,
) -> Dict[str, Any]:
"""
Generic collate function for audio batches where each sample is a dict
containing at least `waveform_key` with shape [1, T] or [T].
Pads or truncates waveforms across the batch, and returns a dict that:
- always includes waveform_key -> Tensor [B, 1, T']
- includes other keys aggregated into lists (or stacked if possible)
Parameters
----------
batch:
List of sample dictionaries.
waveform_key:
Key of waveform in sample dict.
mode:
"pad" -> pad shorter waveforms to max length in batch
"truncate" -> truncate longer waveforms to min length in batch
stack_waveforms:
If True, returns waveforms stacked into a single tensor [B, 1, T'].
pad_value:
Value used for padding.
include_keys:
If provided, only these keys will be included in the output (plus waveform_key).
exclude_keys:
If provided, these keys will not be included (except waveform_key is always kept).
Returns
-------
Dict[str, Any]
Collated batch dict.
"""
if len(batch) == 0:
raise ValueError("Empty batch passed to collate_audio_batch")
# 1) Collect waveforms
waveforms = []
for item in batch:
if waveform_key not in item:
raise KeyError(
f"Missing key '{waveform_key}' in batch item: {list(item.keys())}"
)
w = item[waveform_key]
if not torch.is_tensor(w):
raise TypeError(
f"Expected waveform tensor for key '{waveform_key}', got {type(w)}"
)
# Accept [T] or [1, T]
if w.ndim == 1:
w = w.unsqueeze(0)
elif w.ndim != 2:
raise ValueError(
f"Expected waveform with shape [T] or [1, T], got {tuple(w.shape)}"
)
waveforms.append(w)
lengths = [w.shape[-1] for w in waveforms]
# 2) Determine target length
if mode == "pad":
target_len = max(lengths)
elif mode == "truncate":
target_len = min(lengths)
else:
raise ValueError(f"Unknown mode '{mode}' (expected 'pad' or 'truncate')")
# 3) Pad/truncate each waveform
processed = []
for w in waveforms:
cur_len = w.shape[-1]
if cur_len < target_len:
pad_amount = target_len - cur_len
w2 = torch.nn.functional.pad(w, (0, pad_amount), value=pad_value)
processed.append(w2)
elif cur_len > target_len:
processed.append(w[..., :target_len])
else:
processed.append(w)
if stack_waveforms:
waveform_batch = torch.stack(processed, dim=0) # [B, 1, T']
else:
waveform_batch = processed # list of [1, T']
# 4) Decide which other keys to include
all_keys = set(batch[0].keys())
all_keys.add(waveform_key)
if include_keys is not None:
keys_to_collate = set(include_keys) | {waveform_key}
else:
keys_to_collate = set(all_keys)
if exclude_keys is not None:
keys_to_collate -= set(exclude_keys)
keys_to_collate.add(waveform_key) # waveform always kept
# 5) Collate other keys (best effort)
out: Dict[str, Any] = {waveform_key: waveform_batch}
for k in keys_to_collate:
if k == waveform_key:
continue
values = [item.get(k, None) for item in batch]
# If all are tensors of same shape -> stack
if all(torch.is_tensor(v) for v in values):
try:
out[k] = torch.stack(values, dim=0)
continue
except Exception:
# fallback to list if stacking fails
out[k] = values
continue
# If all are numbers (int/float) -> tensor
if all(isinstance(v, (int, float)) for v in values):
out[k] = torch.tensor(values)
continue
# Otherwise -> list
out[k] = values
return out