experiments
/
STT
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Warn if --audio_sample_rate does not match training sample 

In PR #2688, we started specifying the upper frequency limit when computing Mfccs.
This value was computed as half of the --audio_sample_rate value. Despite accepting
a variable sample rate input for the Mfcc computation, the TensorFlow OP only takes
a constant upper frequency limit, so we can't pass a dynamic value computed from each
sample to the op.

This means we lost the ability to transparently train on data with multiple sample
rates. This commit adds a warning message in case a training sample does not match
the --audio_sample_rate flag.
This commit is contained in:
Reuben Morais 2020-02-18 18:13:36 +01:00
parent 44ff4c54b9
commit 1178215423
1 changed files with 5 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
util/feeding.py
View File

@ -30,7 +30,10 @@ def read_csvs(csv_files):
return pandas.concat(sets, join='inner', ignore_index=True) return pandas.concat(sets, join='inner', ignore_index=True)
def samples_to_mfccs(samples, sample_rate, train_phase=False): def samples_to_mfccs(samples, sample_rate, train_phase=False, wav_filename=None):
if train_phase and sample_rate != FLAGS.audio_sample_rate:
tf.print('WARNING: sample rate of file', wav_filename, '(', sample_rate, ') does not match FLAGS.audio_sample_rate. This can lead to incorrect results.')
spectrogram = contrib_audio.audio_spectrogram(samples, spectrogram = contrib_audio.audio_spectrogram(samples,
window_size=Config.audio_window_samples, window_size=Config.audio_window_samples,
stride=Config.audio_step_samples, stride=Config.audio_step_samples,
@ -79,7 +82,7 @@ def samples_to_mfccs(samples, sample_rate, train_phase=False):
def audiofile_to_features(wav_filename, train_phase=False): def audiofile_to_features(wav_filename, train_phase=False):
samples = tf.io.read_file(wav_filename) samples = tf.io.read_file(wav_filename)
decoded = contrib_audio.decode_wav(samples, desired_channels=1) decoded = contrib_audio.decode_wav(samples, desired_channels=1)
features, features_len = samples_to_mfccs(decoded.audio, decoded.sample_rate, train_phase=train_phase) features, features_len = samples_to_mfccs(decoded.audio, decoded.sample_rate, train_phase=train_phase, wav_filename=wav_filename)
if train_phase: if train_phase:
if FLAGS.data_aug_features_multiplicative > 0: if FLAGS.data_aug_features_multiplicative > 0: