No module named 'torchaudio.kaldi_io' (fix)

On torchaudio 2.9, import torchaudio.kaldi_io raises ModuleNotFoundError: No module named 'torchaudio.kaldi_io'. We reproduced it and tested a drop-in replacement with kaldiio.

1. The error

A one-utterance feature archive (written with kaldiio.save_ark), read the old way:

import torchaudio
import torchaudio.kaldi_io
for key, mat in torchaudio.kaldi_io.read_mat_ark("feats.ark"):
    print(torchaudio.__version__, key, mat.shape, mat.dtype)

torchaudio 2.9.0:

Traceback (most recent call last):
  File "b1.py", line 2, in <module>
    import torchaudio.kaldi_io
ModuleNotFoundError: No module named 'torchaudio.kaldi_io'

torchaudio 2.8.0 with the kaldi_io package installed reads it, after a deprecation warning, a KALDI_ROOT banner from kaldi_io itself, and a read-only-array warning (lines shortened):

b1.py:3: UserWarning: torchaudio.kaldi_io.read_mat_ark has been deprecated. This deprecation is part of a large refactoring effort to transition TorchAudio i...
  for key, mat in torchaudio.kaldi_io.read_mat_ark("feats.ark"):
################################################################################
### WARNING, path does not exist: KALDI_ROOT=/mnt/matylda5/iveselyk/Tools/kaldi-trunk
###          (please add 'export KALDI_ROOT=<your_path>' in your $HOME/.profile)
###          (or run as: KALDI_ROOT=<your_path> python <your_script>.py)
################################################################################

.../site-packages/torchaudio/kaldi_io.py:42: UserWarning: The given NumPy array is not writable, and PyTorch does not support non-writable tensors. This mean...
  yield key, torch.from_numpy(np_arr)
2.8.0+cu128 utt1 torch.Size([2, 3]) torch.float32

2. Fix: read the ark with kaldiio

pip install kaldiio

import torch, kaldiio
for key, mat in kaldiio.load_ark("feats.ark"):
    t = torch.from_numpy(mat.copy())  # copy: kaldiio returns a read-only array
    print(torch.__version__, key, t.shape, t.dtype)

torchaudio 2.9.0 (no torchaudio import needed):

2.9.0+cpu utt1 torch.Size([2, 3]) torch.float32

torch 2.8.0:

2.8.0+cu128 utt1 torch.Size([2, 3]) torch.float32

Same key, shape and dtype as the old read_mat_ark. Without .copy() torch warned "The given NumPy array is not writable" (the old torchaudio wrapper warned the same). For .scp files, kaldiio.load_scp("feats.scp") returns a dict-like reader.

3. Or pin torchaudio 2.8.0

pip install "torch==2.8.0" "torchaudio==2.8.0" kaldi_io

torchaudio.kaldi_io is only a wrapper around the kaldi_io package, so that also has to be installed.

Related: No module named torchaudio.backend · torchaudio sox_effects removed in 2.9 · torchaudio.utils has no attribute download_asset

Written with AI assistance. Measured on 2026-10-24 on Linux with CPython 3.11.17, torch/torchaudio 2.8.0 (PyPI wheels) and 2.9.0 (CPU wheels), kaldiio and kaldi_io from PyPI. Output pasted unedited except shortened paths and trimmed warning lines. PlainTables is our extension; torchaudio and kaldiio are free.