#!/usr/bin/env python3 # -*- coding: utf-8 -*- # # Unit test for FreeDV API resampler functions, from # codec2/unittest/t48_8_short.c - generate a sine wave at 8 KHz, # upsample to 48 kHz, add an interferer, then downsample back to 8 kHz # # You can listen to the output files with: # # aplay -f S16_LE in8.raw # aplay -r 48000 -f S16_LE out48.raw # aplay -f S16_LE out8.raw # # They should sound like clean sine waves # pylint: disable=global-statement, invalid-name, unused-import import os import sys import codec2 import numpy as np import pytest # dig some constants out FDMDV_OS_48 = codec2.api.FDMDV_OS_48 FDMDV_OS_TAPS_48K = codec2.api.FDMDV_OS_TAPS_48K FDMDV_OS_TAPS_48_8K = codec2.api.FDMDV_OS_TAPS_48_8K N8 = 180 # processing buffer size at 8 kHz N48 = N8 * FDMDV_OS_48 # processing buffer size at 48 kHz MEM8 = FDMDV_OS_TAPS_48_8K # 8kHz signal filter memory MEM48 = FDMDV_OS_TAPS_48K # 48kHz signal filter memory FRAMES = 50 # number of frames to test FS8 = 8000 FS48 = 48000 AMP = 16000 # sine wave amplitude FTEST8 = 800 # input test frequency at FS=8kHz FINTER48 = 10000 # interferer frequency at FS=48kHz # Due to the design of these resamplers, the processing buffer (at 8kHz) # must be an integer multiple of oversampling ratio assert N8 % FDMDV_OS_48 == 0 def test_resampler(): """ Test for the codec2 audio resampling routine """ # time indexes, we advance every frame t = 0 t1 = 0 # output files to listen to/evaluate result with open("in8.raw", mode="wb") as fin8: with open("out48.raw", mode="wb") as f48: with open("out8.raw", mode="wb") as fout8: resampler = codec2.resampler() # Generate FRAMES of a sine wave for _ in range(FRAMES): # Primary sine wave, which the down-sampling filter should retain. sine_in8k = ( AMP * np.cos(2 * np.pi * np.arange(t, t + N8) * FTEST8 / FS8) ).astype(np.int16) t += N8 sine_in8k.tofile(fin8) sine_out48k = resampler.resample8_to_48(sine_in8k) sine_out48k.tofile(f48) # Add an interfering sine wave, which the down-sampling filter should (mostly) remove sine_in48k = ( sine_out48k + (AMP / 2) * np.cos(2 * np.pi * np.arange(t1, t1 + N48) * FINTER48 / FS48) ).astype(np.int16) t1 += N48 sine_out8k = resampler.resample48_to_8(sine_in48k) sine_out8k.tofile(fout8) # os.unlink("out48.raw") # Automated test evaluation -------------------------------------------- # The input and output signals will not be time aligned due to the filter # delays, so compare the magnitude spectrum # Read the raw audio files in8k = np.fromfile("in8.raw", dtype=np.int16) out8k = np.fromfile("out8.raw", dtype=np.int16) assert len(in8k) == len(out8k) # os.unlink("in8.raw") # os.unlink("out8.raw") # Apply hanning filter to raw input data samples h = np.hanning(len(in8k)) S1 = np.abs(np.fft.fft(in8k * h)) S2 = np.abs(np.fft.fft(out8k * h)) # Calculate the ratio between signal and noise (error energy). error = S1 - S2 error_energy = np.dot(error, error) ratio = error_energy / np.dot(S1, S1) ratio_dB = 10 * np.log10(ratio) # Establish -40.0 as the noise ratio ceiling threshdB = -40.0 print(f"ratio_dB: {ratio_dB:4.2}" % (ratio_dB)) assert ratio_dB < threshdB if __name__ == "__main__": # Run pytest with the current script as the filename. ecode = pytest.main(["-v", sys.argv[0]]) if ecode == 0: print("PASS") else: print("FAIL")