John Meade johnmeade

## asr.py
"""
Multi-language ASR using Huggingface transformer models.

Python dependencies:
    pip install transformers==4.5.0 librosa soundfile torch
"""

from typing import NamedTuple
from functools import lru_cache

## aeneas_word_force_align.py
"""
Note: Aeneas is based on TTS and DTW, which is not ideal for word-level alignment.
However, it is easy to install and works quite well, so it is still very useful.
This gist just lazily writes files to "/tmp" for demonstration purposes.

System and Python dependencies (Ubuntu):
    sudo apt-get install python-dev espeak espeak-data libespeak1 libespeak-dev ffmpeg
    pip install numpy textgrid
    pip install aeneas

## speech_noise.py
"""
Estimate background noise power level of a speech waveform.
Requires some non-speech regions in the wave.

Requirements:
    pip install numpy librosa soundfile webrtcvad

MIT License John Meade 2021
"""

## snr.py
import numpy as np

def wada_snr(wav):
    # Direct blind estimation of the SNR of a speech signal.
    #
    # Paper on WADA SNR:
    #   http://www.cs.cmu.edu/~robust/Papers/KimSternIS08.pdf
    #
    # This function was adapted from this matlab code:
    #   https://labrosa.ee.columbia.edu/projects/snreval/#9

## pool_data_loader.py
'''
A multiprocessing Pool drop-in replacement for the pytorch
DataLoader class. Built to work around an apparent bug in
the default pytorch DataLoader, in which it hangs indefinitely.
It is possible to reach a sustained 95-100% GPU usage (as
reported by `nvidia-smi`) using this implementation.

Requirements:
    pip install filelock

## dist_sampler.py
'''
Tools for sampling from arbitrary probability densities.

Requirements:
    pip install scipy numpy

John Meade 2019
MIT license
'''
	"""
	Multi-language ASR using Huggingface transformer models.

	Python dependencies:
	pip install transformers==4.5.0 librosa soundfile torch
	"""

	from typing import NamedTuple
	from functools import lru_cache
	"""
	Note: Aeneas is based on TTS and DTW, which is not ideal for word-level alignment.
	However, it is easy to install and works quite well, so it is still very useful.
	This gist just lazily writes files to "/tmp" for demonstration purposes.

	System and Python dependencies (Ubuntu):
	sudo apt-get install python-dev espeak espeak-data libespeak1 libespeak-dev ffmpeg
	pip install numpy textgrid
	pip install aeneas
	"""
	Estimate background noise power level of a speech waveform.
	Requires some non-speech regions in the wave.

	Requirements:
	pip install numpy librosa soundfile webrtcvad

	MIT License John Meade 2021
	"""
	import numpy as np

	def wada_snr(wav):
	# Direct blind estimation of the SNR of a speech signal.
	#
	# Paper on WADA SNR:
	# http://www.cs.cmu.edu/~robust/Papers/KimSternIS08.pdf
	#
	# This function was adapted from this matlab code:
	# https://labrosa.ee.columbia.edu/projects/snreval/#9
	'''
	A multiprocessing Pool drop-in replacement for the pytorch
	DataLoader class. Built to work around an apparent bug in
	the default pytorch DataLoader, in which it hangs indefinitely.
	It is possible to reach a sustained 95-100% GPU usage (as
	reported by `nvidia-smi`) using this implementation.

	Requirements:
	pip install filelock
	'''
	Tools for sampling from arbitrary probability densities.

	Requirements:
	pip install scipy numpy

	John Meade 2019
	MIT license
	'''