No module named 'torchaudio.io' (fix)

from torchaudio.io import StreamReader raises ModuleNotFoundError: No module named 'torchaudio.io' on torchaudio 2.9 and newer (or module 'torchaudio' has no attribute 'io'). We tested the TorchCodec replacements for reading and writing, and the pin-2.8 route with its FFmpeg catch.

1. The error

import torchaudio
print("torchaudio", torchaudio.__version__)
from torchaudio.io import StreamReader

torchaudio 2.9.0:

torchaudio 2.9.0+cpu
Traceback (most recent call last):
  File "b1.py", line 3, in <module>
    from torchaudio.io import StreamReader
ModuleNotFoundError: No module named 'torchaudio.io'

torchaudio 2.8.0 imports it fine:

torchaudio 2.8.0+cu128

import torchaudio; torchaudio.io gives the AttributeError variant of the same thing. The torchaudio.io streaming API (StreamReader, StreamWriter) was removed in 2.9; TorchCodec replaces it.

2. Fix: decode with torchcodec AudioDecoder

torchcodec must match your torch (0.8/0.9 for torch 2.9) and needs FFmpeg 4–8 shared libraries; see Could not load libtorchcodec if it won't import. torch 2.9.0, torchcodec 0.9.1, FFmpeg 7, on a 1-second 16 kHz mono WAV:

from torchcodec.decoders import AudioDecoder
dec = AudioDecoder("tone.wav")
print(dec.metadata.sample_rate, dec.metadata.num_channels, dec.metadata.duration_seconds)
samples = dec.get_all_samples()
print(samples.data.shape, samples.sample_rate)

16000 1 1.0
torch.Size([1, 16000]) 16000

samples.data is a float32 (channels, samples) tensor, the same layout torchaudio.load returns. For chunked reading, dec.get_samples_played_in_range(start, stop) takes seconds.

3. Fix: encode with torchcodec AudioEncoder (replaces StreamWriter)

import torch
from torchcodec.encoders import AudioEncoder
t = torch.linspace(0, 1, 16000)
wave = (0.3 * torch.sin(2 * torch.pi * 440 * t)).unsqueeze(0)   # (channels, samples)
AudioEncoder(wave, sample_rate=16000).to_file("out.flac")
from torchcodec.decoders import AudioDecoder
print(AudioDecoder("out.flac").get_all_samples().data.shape)

torch.Size([1, 16000])

4. Or pin torchaudio 2.8.0 — but mind the FFmpeg version

pip install "torch==2.8.0" "torchaudio==2.8.0" brings torchaudio.io back. Actually reading needs FFmpeg too, and 2.8.0 only looks for FFmpeg 4, 5 or 6:

from torchaudio.io import StreamReader
r = StreamReader("tone.wav")
r.add_basic_audio_stream(frames_per_chunk=16000)
for (chunk,) in r.stream():
    print(chunk.shape)

With no FFmpeg, and again with FFmpeg 7 libraries on LD_LIBRARY_PATH, we got the same error:

Traceback (most recent call last):
  ...
ImportError: Failed to intialize FFmpeg extension. Tried versions: ['6', '5', '4', '']. Enable DEBUG logging to see more details about the error.

So on the 2.8 route install FFmpeg 6 or older (e.g. conda install -c conda-forge "ffmpeg<7"). TorchCodec accepts 4–8. Full walkthrough: Failed to intialize FFmpeg extension.

Related: No module named torchaudio.backend · torchaudio has no attribute info · torchaudio has no attribute AudioMetaData · torchaudio sox_effects: libsox.so not found

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), torchcodec 0.9.1, conda-forge FFmpeg 7. Output pasted unedited except shortened site-packages paths and cut tracebacks. PlainTables is our extension; torchaudio and torchcodec are free.