Example code, librosa.util.exceptions.ParameterError: Audio data must be floating-point

[vaughantnrc]

Hi everyone, I searched for exception "Audio data must be floating-point" but found no results.

I'm trying to run code that is more or less identical to this: https://github.com/NVIDIA/NeMo/tree/main/examples/speaker_tasks/diarization

Except with a .wav file. When I do so I get the following exception:

Traceback (most recent call last):
  File "**********\venv\lib\site-packages\nemo\collections\asr\parts\utils\decoder_timestamps_utils.py", line 661, in run_ASR_BPE_CTC
    hyp, greedy_predictions_list, log_prob = get_wer_feat_logit(
  File "**********\venv\lib\site-packages\nemo\collections\asr\parts\utils\decoder_timestamps_utils.py", line 221, in get_wer_feat_logit
    asr.read_audio_file_and_return(audio_file_path, delay, model_stride_in_secs)
  File "**********\venv\lib\site-packages\nemo\collections\asr\parts\utils\decoder_timestamps_utils.py", line 264, in read_audio_file_and_return
    samples = get_samples(audio_filepath)
  File "**********\venv\lib\site-packages\nemo\collections\asr\parts\utils\decoder_timestamps_utils.py", line 235, in get_samples
    samples = librosa.core.resample(samples, sample_rate, target_sr)
  File "**********\venv\lib\site-packages\librosa\util\decorators.py", line 104, in inner_f
    return f(**kwargs)
  File "**********\venv\lib\site-packages\librosa\core\audio.py", line 606, in resample
    util.valid_audio(y, mono=False)
  File "**********\venv\lib\site-packages\librosa\util\decorators.py", line 88, in inner_f
    return f(*args, **kwargs)
  File "**********\venv\lib\site-packages\librosa\util\utils.py", line 275, in valid_audio
    raise ParameterError("Audio data must be floating-point")
librosa.util.exceptions.ParameterError: Audio data must be floating-point

When I look at NeMo's source code, I see that the audio is explicitly loaded with type int16: https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/asr/parts/utils/decoder_timestamps_utils.py#L231

What confuses me is that if I look at the history of librosa, I see that the check for integer types has existed for years: https://github.com/librosa/librosa/blob/main/librosa/core/audio.py

A quick glance at the functions in the stack did not turn up any conversion from int16 to float.

So my question is... Is this a bug? The code looks wrong to me, but I have trouble imagining that this issue has existed without being detected, especially when there is an example (linked above)?

In case it's pertinent:

• I'm using NeMo 1.11.0 pip install nemo-toolkit[asr]==1.11.0
• I appear to be on the latest version of librosa, which is version 0.9.2
• I'm on Windows (though based on the observations/code above I do not see any obvious link operating system)

[titu1994]

@nithinraok possibly a bug?

[nithinraok]

We don't have test cases to test diarization on samples not sampled at 16kHz, that was why it wasn't caught. Thanks for reporting. May be @tango4j can explain why this was done to load samples explicitly as int16.

[tango4j]

@vaughantnrc Can you share the specification of the audio file that caused this error? type of wave file, sampling rate, bit depth etc. Preferably sox --i abc.wav output. We may need to change the get_samples function for better generalizability.

[vaughantnrc]

Thank you for looking into this. This is what I get with sox:

Input File     : 'sim.wav'
Channels       : 1
Sample Rate    : 48000
Precision      : 16-bit
Duration       : 00:02:40.30 = 7694400 samples ~ 12022.5 CDDA sectors
File Size      : 15.4M
Bit Rate       : 768k
Sample Encoding: 16-bit Signed Integer PCM

[tango4j]

@vaughantnrc Sorry for the delayed response. We are working on a PullRequest for fixing this issue. Feel free comment on this PR if you think it is necessary.