fix(asr)cache-aware streaming: mono audio and variables UnboundLocalError

[YCmove]

What does this PR do ?

• Several variables are not properly initialized in the single-input audio case.
• Stereo audio is not converted to mono, causing the downstream error.

Collection: asr

Changelog

• Initialize variables correctly: all_streaming_tran, all_offline_tran, all_refs_text, and start_time
• Add an error message when output_path is specified without dataset_manifest.
• Always read audio_file in mono.

Usage

Feed the mono or stereo audio_file to examples/asr/asr_cache_aware_streaming/speech_to_text_cache_aware_streaming_infer.py.

python speech_to_text_cache_aware_streaming_infer.py model_path={path to nemotron-3.5-asr-streaming-0.6b.nemo} audio_file={path to any mono or stereo audio in wav or flac} compare_vs_offline={true/flase} amp={true/flase} debug_mode={true/flase} target_lang={target language}

Test Environments: RTX 3090, CUDA Version: 13.2, Ubuntu 26.04, torch 2.12.0+cu132

Verification

Following tests are passed without warning:

uv run pytest tests/collections/asr/utils/test_streaming_utils.py
uv run pytest tests/collections/asr/test_preprocessing_segment.py
uv run pytest tests/collections/asr/inference/test_cache_aware_pipeline_builder.py

GitHub Actions CI

The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.

Trusted PRs run automatically through copy-pr-bot. For an untrusted PR, a maintainer can trigger CI by commenting
/ok to test <head-sha>; repeat this after a new push if the PR remains untrusted.

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
• Reviewer: Does the PR have correct import guards for all optional libraries?

PR Type:

• New Feature
• Bugfix
• Documentation

Who can review?

@nithinraok or the previous author @artbataev
Anyone in the community is free to review the PR once the checks have passed.
Contributor guidelines contains specific people who can review PRs to various areas.

Additional Information

None

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[nithinraok]

revert this, we used soundfile as its multiple times faster than librosa. What is the motivation for this change?

[YCmove]

This is more about the stereo audio not being handled correctly.

1. librosa resamples the stereo audio in the wrong dimension
   resample takes input format as np.ndarray [shape=(n,) or shape=(2, n)], whereas soundfile reads stereo as a two-dimensional array in (n, 2). This works fine for mono because soundfile reads mono as a one-dimensional array (n,) by default.
   samples = samples.transpose() seems to be unnecessary, the transpose of (n,) is still (n,)

2. This causes errors downstream in streaming_utils.py,

   1. audio_signal_len=2 instead of n
   2. incorrect format for audio_signal to pass into the preprocessor: (1, 2, n) instead of (batch, n)

I roll back to the original sf.SoundFile but add code to convert stereo to mono.