torchaudio.load slows down with increasing frame_offset

🐛 Describe the bug

The execution time of torchaudio.load increases when increasing frame_offset. This can be observed in particular when loading a WAV file sequentially in chunks.

I assume that here load reads the whole file up until frame_offset and then reads the requested chunk, which is inefficient. While this might be compilcated to fix for compressed file formats, for WAV files this could be solved by simply seeking into the file and reading the file starting at frame_offset.

Here's a minimal example to reproduce the issue:

from time import time
import torchaudio


WAV_FILE = "path/to/WAV/file.wav"
CHUNK_SIZE = 2**16


def load_benchmark():
    try:
        info = torchaudio.info(str(WAV_FILE))
    except RuntimeError:
        print(WAV_FILE)
        raise
    length = info.num_frames

    for offset in range(0, length, CHUNK_SIZE):
        start = time()
        _, _ = torchaudio.load(WAV_FILE, frame_offset=offset, num_frames=CHUNK_SIZE)
        end = time()
        print(f"Loading chunk at offset {offset} took {(end - start) * 1000}ms.")


if __name__ == '__main__':
    load_benchmark()

which outputs something like

Loading chunk at offset 0 took 8.82411003112793ms.
Loading chunk at offset 65536 took 5.040168762207031ms.
[...]
Loading chunk at offset 9437184 took 72.36099243164062ms.
Loading chunk at offset 9502720 took 72.80373573303223ms.

Versions

PyTorch version: 1.13.0
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: macOS 12.5 (x86_64)
GCC version: Could not collect
Clang version: 14.0.0 (clang-1400.0.29.102)
CMake version: version 3.20.5
Libc version: N/A

Python version: 3.9.15 (main, Nov 24 2022, 08:29:02)  [Clang 14.0.6 ] (64-bit runtime)
Python platform: macOS-10.16-x86_64-i386-64bit
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

Versions of relevant libraries:
[pip3] mypy==0.991
[pip3] mypy-extensions==0.4.3
[pip3] numpy==1.23.5
[pip3] torch==1.13.0
[pip3] torchaudio==0.13.0
[pip3] torchmetrics==0.11.0
[conda] numpy                     1.23.5                   pypi_0    pypi
[conda] torch                     1.13.0                   pypi_0    pypi
[conda] torchaudio                0.13.0                   pypi_0    pypi
[conda] torchmetrics              0.11.0                   pypi_0    pypi

Start at the torchaudio.load entry point and reproduce the behavior with the provided sequential WAV-file benchmark using increasing frame_offset values. Trace how WAV loading handles frame_offset and compare the elapsed times across chunks; done means WAV chunks can be loaded without time increasing with the offset, while preserving the requested frame range.