MLECO-1253 Adding ASR sample application using the PyArmNN api
Change-Id: I450b23800ca316a5bfd4608c8559cf4f11271c21
Signed-off-by: Éanna Ó Catháin <eanna.ocathain@arm.com>
diff --git a/python/pyarmnn/examples/speech_recognition/tests/conftest.py b/python/pyarmnn/examples/speech_recognition/tests/conftest.py
new file mode 100644
index 0000000..730c291
--- /dev/null
+++ b/python/pyarmnn/examples/speech_recognition/tests/conftest.py
@@ -0,0 +1,34 @@
+# Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
+# SPDX-License-Identifier: MIT
+
+import os
+import ntpath
+
+import urllib.request
+
+import pytest
+
+script_dir = os.path.dirname(__file__)
+
+@pytest.fixture(scope="session")
+def test_data_folder(request):
+ """
+ This fixture returns path to folder with shared test resources among all tests
+ """
+
+ data_dir = os.path.join(script_dir, "testdata")
+
+ if not os.path.exists(data_dir):
+ os.mkdir(data_dir)
+
+ files_to_download = ["https://raw.githubusercontent.com/Azure-Samples/cognitive-services-speech-sdk/master"
+ "/sampledata/audiofiles/myVoiceIsMyPassportVerifyMe04.wav"]
+
+ for file in files_to_download:
+ path, filename = ntpath.split(file)
+ file_path = os.path.join(script_dir, "testdata", filename)
+ if not os.path.exists(file_path):
+ print("\nDownloading test file: " + file_path + "\n")
+ urllib.request.urlretrieve(file, file_path)
+
+ return data_dir
diff --git a/python/pyarmnn/examples/speech_recognition/tests/context.py b/python/pyarmnn/examples/speech_recognition/tests/context.py
new file mode 100644
index 0000000..a810010
--- /dev/null
+++ b/python/pyarmnn/examples/speech_recognition/tests/context.py
@@ -0,0 +1,13 @@
+# Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
+# SPDX-License-Identifier: MIT
+
+import os
+import sys
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..', '..', 'common'))
+import utils as common_utils
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..'))
+import audio_capture
+import audio_utils
+import preprocess
diff --git a/python/pyarmnn/examples/speech_recognition/tests/test_audio_file.py b/python/pyarmnn/examples/speech_recognition/tests/test_audio_file.py
new file mode 100644
index 0000000..281d0df
--- /dev/null
+++ b/python/pyarmnn/examples/speech_recognition/tests/test_audio_file.py
@@ -0,0 +1,17 @@
+# Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
+# SPDX-License-Identifier: MIT
+
+import os
+
+import numpy as np
+
+from context import audio_capture
+
+
+def test_audio_file(test_data_folder):
+ audio_file = os.path.join(test_data_folder, "myVoiceIsMyPassportVerifyMe04.wav")
+ capture = audio_capture.AudioCapture(audio_capture.ModelParams(""))
+ buffer = capture.from_audio_file(audio_file)
+ audio_data = next(buffer)
+ assert audio_data.shape == (47712,)
+ assert audio_data.dtype == np.float32
diff --git a/python/pyarmnn/examples/speech_recognition/tests/test_decoder.py b/python/pyarmnn/examples/speech_recognition/tests/test_decoder.py
new file mode 100644
index 0000000..3b99e65
--- /dev/null
+++ b/python/pyarmnn/examples/speech_recognition/tests/test_decoder.py
@@ -0,0 +1,28 @@
+# Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
+# SPDX-License-Identifier: MIT
+
+import os
+
+import numpy as np
+
+from context import common_utils
+from context import audio_utils
+
+
+def test_labels(test_data_folder):
+ labels_file = os.path.join(test_data_folder, "wav2letter_labels.txt")
+ labels = common_utils.dict_labels(labels_file)
+ assert len(labels) == 29
+ assert labels[26] == "\'"
+ assert labels[27] == r" "
+ assert labels[28] == "$"
+
+
+def test_decoder(test_data_folder):
+ labels_file = os.path.join(test_data_folder, "wav2letter_labels.txt")
+ labels = common_utils.dict_labels(labels_file)
+
+ output_tensor = os.path.join(test_data_folder, "inf_out.npy")
+ encoded = np.load(output_tensor)
+ decoded_text = audio_utils.decode(encoded, labels)
+ assert decoded_text == "and he walkd immediately out of the apartiment by anothe"
diff --git a/python/pyarmnn/examples/speech_recognition/tests/test_mfcc.py b/python/pyarmnn/examples/speech_recognition/tests/test_mfcc.py
new file mode 100644
index 0000000..d692ab5
--- /dev/null
+++ b/python/pyarmnn/examples/speech_recognition/tests/test_mfcc.py
@@ -0,0 +1,286 @@
+# Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
+# SPDX-License-Identifier: MIT
+
+import numpy as np
+
+from context import preprocess
+
+test_wav = [
+ -3,0,1,-1,2,3,-2,2,
+ 1,-2,0,3,-1,8,3,2,
+ -1,-1,2,7,3,5,6,6,
+ 6,12,5,6,3,3,5,4,
+ 4,6,7,7,7,3,7,2,
+ 8,4,4,2,-4,-1,-1,-4,
+ 2,1,-1,-4,0,-7,-6,-2,
+ -5,1,-5,-1,-7,-3,-3,-7,
+ 0,-3,3,-5,0,1,-2,-2,
+ -3,-3,-7,-3,-2,-6,-5,-8,
+ -2,-8,4,-9,-4,-9,-5,-5,
+ -3,-9,-3,-9,-1,-7,-4,1,
+ -3,2,-8,-4,-4,-5,1,-3,
+ -1,0,-1,-2,-3,-2,-4,-1,
+ 1,-1,3,0,3,2,0,0,
+ 0,-3,1,1,0,8,3,4,
+ 1,5,6,4,7,3,3,0,
+ 3,6,7,6,4,5,9,9,
+ 5,5,8,1,6,9,6,6,
+ 7,1,8,1,5,0,5,5,
+ 0,3,2,7,2,-3,3,0,
+ 3,0,0,0,2,0,-1,-1,
+ -2,-3,-8,0,1,0,-3,-3,
+ -3,-2,-3,-3,-4,-6,-2,-8,
+ -9,-4,-1,-5,-3,-3,-4,-3,
+ -6,3,0,-1,-2,-9,-4,-2,
+ 2,-1,3,-5,-5,-2,0,-2,
+ 0,-1,-3,1,-2,9,4,5,
+ 2,2,1,0,-6,-2,0,0,
+ 0,-1,4,-4,3,-7,-1,5,
+ -6,-1,-5,4,3,9,-2,1,
+ 3,0,0,-2,1,2,1,1,
+ 0,3,2,-1,3,-3,7,0,
+ 0,3,2,2,-2,3,-2,2,
+ -3,4,-1,-1,-5,-1,-3,-2,
+ 1,-1,3,2,4,1,2,-2,
+ 0,2,7,0,8,-3,6,-3,
+ 6,1,2,-3,-1,-1,-1,1,
+ -2,2,1,2,0,-2,3,-2,
+ 3,-2,1,0,-3,-1,-2,-4,
+ -6,-5,-8,-1,-4,0,-3,-1,
+ -1,-1,0,-2,-3,-7,-1,0,
+ 1,5,0,5,1,1,-3,0,
+ -6,3,-8,4,-8,6,-6,1,
+ -6,-2,-5,-6,0,-5,4,-1,
+ 4,-2,1,2,1,0,-2,0,
+ 0,2,-2,2,-5,2,0,-2,
+ 1,-2,0,5,1,0,1,5,
+ 0,8,3,2,2,0,5,-2,
+ 3,1,0,1,0,-2,-1,-3,
+ 1,-1,3,0,3,0,-2,-1,
+ -4,-4,-4,-1,-4,-4,-3,-6,
+ -3,-7,-3,-1,-2,0,-5,-4,
+ -7,-3,-2,-2,1,2,2,8,
+ 5,4,2,4,3,5,0,3,
+ 3,6,4,2,2,-2,4,-2,
+ 3,3,2,1,1,4,-5,2,
+ -3,0,-1,1,-2,2,5,1,
+ 4,2,3,1,-1,1,0,6,
+ 0,-2,-1,1,-1,2,-5,-1,
+ -5,-1,-6,-3,-3,2,4,0,
+ -1,-5,3,-4,-1,-3,-4,1,
+ -4,1,-1,-1,0,-5,-4,-2,
+ -1,-1,-3,-7,-3,-3,4,4,
+]
+
+def test_mel_scale_function_with_htk_true():
+ samp_freq = 16000
+ frame_len_ms = 32
+ frame_len_samples = samp_freq * frame_len_ms * 0.001
+ num_mfcc_feats = 13
+ num_fbank_bins = 128
+ mel_lo_freq = 0
+ mil_hi_freq = 8000
+ use_htk = False
+ n_FFT = 512
+
+ mfcc_params = preprocess.MFCCParams(samp_freq, num_fbank_bins, mel_lo_freq, mil_hi_freq, num_mfcc_feats,
+ frame_len_samples, use_htk, n_FFT)
+
+ mfcc_inst = preprocess.MFCC(mfcc_params)
+
+ mel = mfcc_inst.mel_scale(16, True)
+
+ assert np.isclose(mel, 25.470010570730597)
+
+
+def test_mel_scale_function_with_htk_false():
+ samp_freq = 16000
+ frame_len_ms = 32
+ frame_len_samples = samp_freq * frame_len_ms * 0.001
+ num_mfcc_feats = 13
+ num_fbank_bins = 128
+ mel_lo_freq = 0
+ mil_hi_freq = 8000
+ use_htk = False
+ n_FFT = 512
+
+ mfcc_params = preprocess.MFCCParams(samp_freq, num_fbank_bins, mel_lo_freq, mil_hi_freq, num_mfcc_feats,
+ frame_len_samples, use_htk, n_FFT)
+
+ mfcc_inst = preprocess.MFCC(mfcc_params)
+
+ mel = mfcc_inst.mel_scale(16, False)
+
+ assert np.isclose(mel, 0.24)
+
+
+def test_inverse_mel_scale_function_with_htk_true():
+ samp_freq = 16000
+ frame_len_ms = 32
+ frame_len_samples = samp_freq * frame_len_ms * 0.001
+ num_mfcc_feats = 13
+ num_fbank_bins = 128
+ mel_lo_freq = 0
+ mil_hi_freq = 8000
+ use_htk = False
+ n_FFT = 512
+
+ mfcc_params = preprocess.MFCCParams(samp_freq, num_fbank_bins, mel_lo_freq, mil_hi_freq, num_mfcc_feats,
+ frame_len_samples, use_htk, n_FFT)
+
+ mfcc_inst = preprocess.MFCC(mfcc_params)
+
+ mel = mfcc_inst.inv_mel_scale(16, True)
+
+ assert np.isclose(mel, 10.008767240008943)
+
+
+def test_inverse_mel_scale_function_with_htk_false():
+ samp_freq = 16000
+ frame_len_ms = 32
+ frame_len_samples = samp_freq * frame_len_ms * 0.001
+ num_mfcc_feats = 13
+ num_fbank_bins = 128
+ mel_lo_freq = 0
+ mil_hi_freq = 8000
+ use_htk = False
+ n_FFT = 512
+
+ mfcc_params = preprocess.MFCCParams(samp_freq, num_fbank_bins, mel_lo_freq, mil_hi_freq, num_mfcc_feats,
+ frame_len_samples, use_htk, n_FFT)
+
+ mfcc_inst = preprocess.MFCC(mfcc_params)
+
+ mel = mfcc_inst.inv_mel_scale(16, False)
+
+ assert np.isclose(mel, 1071.170287494467)
+
+
+def test_create_mel_filter_bank():
+ samp_freq = 16000
+ frame_len_ms = 32
+ frame_len_samples = samp_freq * frame_len_ms * 0.001
+ num_mfcc_feats = 13
+ num_fbank_bins = 128
+ mel_lo_freq = 0
+ mil_hi_freq = 8000
+ use_htk = False
+ n_FFT = 512
+
+ mfcc_params = preprocess.MFCCParams(samp_freq, num_fbank_bins, mel_lo_freq, mil_hi_freq, num_mfcc_feats,
+ frame_len_samples, use_htk, n_FFT)
+
+ mfcc_inst = preprocess.MFCC(mfcc_params)
+
+ mel_filter_bank = mfcc_inst.create_mel_filter_bank()
+
+ assert len(mel_filter_bank) == 128
+
+ assert str(mel_filter_bank[0]) == "[0.02837754]"
+ assert str(mel_filter_bank[1]) == "[0.01438901 0.01398853]"
+ assert str(mel_filter_bank[2]) == "[0.02877802]"
+ assert str(mel_filter_bank[3]) == "[0.04236608]"
+ assert str(mel_filter_bank[4]) == "[0.00040047 0.02797707]"
+ assert str(mel_filter_bank[5]) == "[0.01478948 0.01358806]"
+ assert str(mel_filter_bank[50]) == "[0.03298853]"
+ assert str(mel_filter_bank[100]) == "[0.00260166 0.00588759 0.00914814 0.00798015 0.00476919 0.00158245]"
+
+
+def test_mfcc_compute():
+ samp_freq = 16000
+ frame_len_ms = 32
+ frame_len_samples = samp_freq * frame_len_ms * 0.001
+ num_mfcc_feats = 13
+ num_fbank_bins = 128
+ mel_lo_freq = 0
+ mil_hi_freq = 8000
+ use_htk = False
+ n_FFT = 512
+
+ audio_data = np.array(test_wav) / (2 ** 15)
+
+ mfcc_params = preprocess.MFCCParams(samp_freq, num_fbank_bins, mel_lo_freq, mil_hi_freq, num_mfcc_feats,
+ frame_len_samples, use_htk, n_FFT)
+ mfcc_inst = preprocess.MFCC(mfcc_params)
+ mfcc_feats = mfcc_inst.mfcc_compute(audio_data)
+
+ assert np.isclose((mfcc_feats[0]), -834.9656973095651)
+ assert np.isclose((mfcc_feats[1]), 21.026915475076322)
+ assert np.isclose((mfcc_feats[2]), 18.628541708201688)
+ assert np.isclose((mfcc_feats[3]), 7.341153529494758)
+ assert np.isclose((mfcc_feats[4]), 18.907974386153214)
+ assert np.isclose((mfcc_feats[5]), -5.360387487466194)
+ assert np.isclose((mfcc_feats[6]), 6.523572638527085)
+ assert np.isclose((mfcc_feats[7]), -11.270643644983316)
+ assert np.isclose((mfcc_feats[8]), 8.375177203773777)
+ assert np.isclose((mfcc_feats[9]), 12.06721844362991)
+ assert np.isclose((mfcc_feats[10]), 8.30815892468875)
+ assert np.isclose((mfcc_feats[11]), -13.499911910889917)
+ assert np.isclose((mfcc_feats[12]), -18.176121251436165)
+
+
+def test_sliding_window_for_small_num_samples():
+ samp_freq = 16000
+ frame_len_ms = 32
+ frame_len_samples = samp_freq * frame_len_ms * 0.001
+ num_mfcc_feats = 13
+ mode_input_size = 9
+ stride = 160
+ num_fbank_bins = 128
+ mel_lo_freq = 0
+ mil_hi_freq = 8000
+ use_htk = False
+ n_FFT = 512
+
+ audio_data = np.array(test_wav) / (2 ** 15)
+
+ full_audio_data = np.tile(audio_data, 9)
+
+ mfcc_params = preprocess.MFCCParams(samp_freq, num_fbank_bins, mel_lo_freq, mil_hi_freq, num_mfcc_feats,
+ frame_len_samples, use_htk, n_FFT)
+ mfcc_inst = preprocess.MFCC(mfcc_params)
+ preprocessor = preprocess.Preprocessor(mfcc_inst, mode_input_size, stride)
+
+ input_tensor = preprocessor.extract_features(full_audio_data)
+
+ assert np.isclose(input_tensor[0][0], -3.4660944830426454)
+ assert np.isclose(input_tensor[0][1], 0.3587718932127629)
+ assert np.isclose(input_tensor[0][2], 0.3480551325669172)
+ assert np.isclose(input_tensor[0][3], 0.2976191917228921)
+ assert np.isclose(input_tensor[0][4], 0.3493037340849936)
+ assert np.isclose(input_tensor[0][5], 0.2408643285767937)
+ assert np.isclose(input_tensor[0][6], 0.2939659585037282)
+ assert np.isclose(input_tensor[0][7], 0.2144552669573928)
+ assert np.isclose(input_tensor[0][8], 0.302239565899944)
+ assert np.isclose(input_tensor[0][9], 0.3187368787077345)
+ assert np.isclose(input_tensor[0][10], 0.3019401051295793)
+ assert np.isclose(input_tensor[0][11], 0.20449412797602678)
+
+ assert np.isclose(input_tensor[0][38], -0.18751440767749533)
+
+
+def test_sliding_window_for_wav_2_letter_sized_input():
+ samp_freq = 16000
+ frame_len_ms = 32
+ frame_len_samples = samp_freq * frame_len_ms * 0.001
+ num_mfcc_feats = 13
+ mode_input_size = 296
+ stride = 160
+ num_fbank_bins = 128
+ mel_lo_freq = 0
+ mil_hi_freq = 8000
+ use_htk = False
+ n_FFT = 512
+
+ audio_data = np.zeros(47712, dtype=int)
+
+ mfcc_params = preprocess.MFCCParams(samp_freq, num_fbank_bins, mel_lo_freq, mil_hi_freq, num_mfcc_feats,
+ frame_len_samples, use_htk, n_FFT)
+
+ mfcc_inst = preprocess.MFCC(mfcc_params)
+ preprocessor = preprocess.Preprocessor(mfcc_inst, mode_input_size, stride)
+
+ input_tensor = preprocessor.extract_features(audio_data)
+
+ assert len(input_tensor[0]) == 39
+ assert len(input_tensor) == 296
diff --git a/python/pyarmnn/examples/speech_recognition/tests/testdata/inf_out.npy b/python/pyarmnn/examples/speech_recognition/tests/testdata/inf_out.npy
new file mode 100644
index 0000000..a6f9ec0
--- /dev/null
+++ b/python/pyarmnn/examples/speech_recognition/tests/testdata/inf_out.npy
Binary files differ
diff --git a/python/pyarmnn/examples/speech_recognition/tests/testdata/wav2letter_labels.txt b/python/pyarmnn/examples/speech_recognition/tests/testdata/wav2letter_labels.txt
new file mode 100644
index 0000000..d7485b7
--- /dev/null
+++ b/python/pyarmnn/examples/speech_recognition/tests/testdata/wav2letter_labels.txt
@@ -0,0 +1,29 @@
+a
+b
+c
+d
+e
+f
+g
+h
+i
+j
+k
+l
+m
+n
+o
+p
+q
+r
+s
+t
+u
+v
+w
+x
+y
+z
+'
+
+$
\ No newline at end of file