Workaround: SpeechBrain torchaudio.info() removed in torchaudio 2.9.x

[bobsummerwill]

Summary

SpeechBrain's dataio.py uses torchaudio.info() and torchaudio.backend.common.AudioMetaData which were removed in torchaudio 2.9.x. This causes import errors and runtime failures when processing audio files.

Error

AttributeError: module 'torchaudio' has no attribute 'info'

Or:

ModuleNotFoundError: No module named 'torchaudio.backend.common'

Root Cause

• torchaudio 2.9.x removed torchaudio.info() function
• torchaudio 2.9.x removed torchaudio.backend.common.AudioMetaData class
• SpeechBrain uses these to get audio file metadata in read_audio_info()

Affected Versions

• torchaudio >= 2.9.0
• SpeechBrain (current versions as of Jan 2026)

Workaround

The install script (Step 10b) patches speechbrain/dataio/dataio.py:

1. Adds AudioMetaDataCompat class to replace the removed AudioMetaData
2. Replaces read_audio_info() to use torchaudio.load() for metadata extraction:

class AudioMetaDataCompat:
    """Compatibility class replacing torchaudio.backend.common.AudioMetaData."""
    def __init__(self, sample_rate, num_frames, num_channels, bits_per_sample=16, encoding="PCM_S"):
        self.sample_rate = sample_rate
        self.num_frames = num_frames
        self.num_channels = num_channels
        self.bits_per_sample = bits_per_sample
        self.encoding = encoding

def read_audio_info(path, backend=None) -> AudioMetaDataCompat:
    if hasattr(torchaudio, 'info'):
        # Old torchaudio version - use original approach
        ...
    else:
        # torchaudio 2.9.x: info() removed, use load() to get metadata
        channels_data, sample_rate = torchaudio.load(path, normalize=False)
        return AudioMetaDataCompat(
            sample_rate=sample_rate,
            num_frames=channels_data.size(1),
            num_channels=channels_data.size(0),
        )

Resolution

This workaround can be removed once SpeechBrain releases a version compatible with torchaudio 2.9.x.

[bobsummerwill-openclaw-bot]

Upstream Status Update (2026-03-06)

SpeechBrain has addressed this on main branch:

• Commit 2427785d (2025-11-22): "Replace torchaudio I/O with soundfile-based audio_io wrapper (#2989)"
• This removes the dependency on torchaudio.info() entirely — same approach as our workaround patch.

However, this fix has NOT shipped in a release yet. Latest release is v1.0.3 (2025-04-07), 7 months before the fix.

Action: Our workaround patches remain necessary until SpeechBrain cuts a new release. Could alternatively pin to SpeechBrain main branch to drop patches sooner (riskier).

Related: speechbrain/speechbrain#3012 (open) tracks the same list_audio_backends removal (#38).

[bobsummerwill-openclaw-bot]

Upstream Fix Available

SpeechBrain v1.0.3 (released March 12, 2025) includes a fix for this.

The current SpeechBrain main branch has read_audio_info() in dataio.py wrapped in a try/except that falls back to _read_audio_info_fallback() using soundfile.info() when torchaudio.info() is unavailable.

Action: Upgrade to SpeechBrain >= 1.0.3 and remove our monkey-patch from install_packages_and_venv.sh (Step 10b).

Closing — upstream fix is released.

[bobsummerwill-openclaw-bot]

Reopening — upstream fix is available but we haven't actually updated and tested yet. Keeping open until the dependency is upgraded and the workaround is confirmed removable.