1/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
2
3Licensed under the Apache License, Version 2.0 (the "License");
4you may not use this file except in compliance with the License.
5You may obtain a copy of the License at
6
7 http://www.apache.org/licenses/LICENSE-2.0
8
9Unless required by applicable law or agreed to in writing, software
10distributed under the License is distributed on an "AS IS" BASIS,
11WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12See the License for the specific language governing permissions and
13limitations under the License.
14==============================================================================*/
15
16// See docs in ../ops/audio_ops.cc
17
18#include "tensorflow/core/framework/op_kernel.h"
19#include "tensorflow/core/framework/register_types.h"
20#include "tensorflow/core/framework/tensor.h"
21#include "tensorflow/core/framework/tensor_shape.h"
22#include "tensorflow/core/framework/types.h"
23#include "tensorflow/core/kernels/mfcc.h"
24#include "tensorflow/core/lib/core/status.h"
25
26namespace tensorflow {
27
28// Create a speech fingerpring from spectrogram data.
29class MfccOp : public OpKernel {
30 public:
31 explicit MfccOp(OpKernelConstruction* context) : OpKernel(context) {
32 OP_REQUIRES_OK(context, context->GetAttr("upper_frequency_limit",
33 &upper_frequency_limit_));
34 OP_REQUIRES_OK(context, context->GetAttr("lower_frequency_limit",
35 &lower_frequency_limit_));
36 OP_REQUIRES_OK(context, context->GetAttr("filterbank_channel_count",
37 &filterbank_channel_count_));
38 OP_REQUIRES_OK(context, context->GetAttr("dct_coefficient_count",
39 &dct_coefficient_count_));
40 }
41
42 void Compute(OpKernelContext* context) override {
43 const Tensor& spectrogram = context->input(0);
44 OP_REQUIRES(context, spectrogram.dims() == 3,
45 errors::InvalidArgument("spectrogram must be 3-dimensional",
46 spectrogram.shape().DebugString()));
47 const Tensor& sample_rate_tensor = context->input(1);
48 OP_REQUIRES(context, TensorShapeUtils::IsScalar(sample_rate_tensor.shape()),
49 errors::InvalidArgument(
50 "Input sample_rate should be a scalar tensor, got ",
51 sample_rate_tensor.shape().DebugString(), " instead."));
52 const int32_t sample_rate = sample_rate_tensor.scalar<int32>()();
53
54 const int spectrogram_channels = spectrogram.dim_size(2);
55 const int spectrogram_samples = spectrogram.dim_size(1);
56 const int audio_channels = spectrogram.dim_size(0);
57
58 Mfcc mfcc;
59 mfcc.set_upper_frequency_limit(upper_frequency_limit_);
60 mfcc.set_lower_frequency_limit(lower_frequency_limit_);
61 mfcc.set_filterbank_channel_count(filterbank_channel_count_);
62 mfcc.set_dct_coefficient_count(dct_coefficient_count_);
63 OP_REQUIRES(context, mfcc.Initialize(spectrogram_channels, sample_rate),
64 errors::InvalidArgument(
65 "Mfcc initialization failed for channel count ",
66 spectrogram_channels, " and sample rate ", sample_rate));
67
68 Tensor* output_tensor = nullptr;
69 OP_REQUIRES_OK(context,
70 context->allocate_output(
71 0,
72 TensorShape({audio_channels, spectrogram_samples,
73 dct_coefficient_count_}),
74 &output_tensor));
75
76 const float* spectrogram_flat = spectrogram.flat<float>().data();
77 float* output_flat = output_tensor->flat<float>().data();
78
79 for (int audio_channel = 0; audio_channel < audio_channels;
80 ++audio_channel) {
81 for (int spectrogram_sample = 0; spectrogram_sample < spectrogram_samples;
82 ++spectrogram_sample) {
83 const float* sample_data =
84 spectrogram_flat +
85 (audio_channel * spectrogram_samples * spectrogram_channels) +
86 (spectrogram_sample * spectrogram_channels);
87 std::vector<double> mfcc_input(sample_data,
88 sample_data + spectrogram_channels);
89 std::vector<double> mfcc_output;
90 mfcc.Compute(mfcc_input, &mfcc_output);
91 DCHECK_EQ(dct_coefficient_count_, mfcc_output.size());
92 float* output_data =
93 output_flat +
94 (audio_channel * spectrogram_samples * dct_coefficient_count_) +
95 (spectrogram_sample * dct_coefficient_count_);
96 for (int i = 0; i < dct_coefficient_count_; ++i) {
97 output_data[i] = mfcc_output[i];
98 }
99 }
100 }
101 }
102
103 private:
104 float upper_frequency_limit_;
105 float lower_frequency_limit_;
106 int32 filterbank_channel_count_;
107 int32 dct_coefficient_count_;
108};
109REGISTER_KERNEL_BUILDER(Name("Mfcc").Device(DEVICE_CPU), MfccOp);
110
111} // namespace tensorflow
112