offline-recognizer-whisper-impl.h
4.3 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
// sherpa-onnx/csrc/offline-recognizer-whisper-impl.h
//
// Copyright (c) 2022-2023 Xiaomi Corporation
#ifndef SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_WHISPER_IMPL_H_
#define SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_WHISPER_IMPL_H_
#include <algorithm>
#include <cmath>
#include <memory>
#include <string>
#include <utility>
#include <vector>
#include "sherpa-onnx/csrc/offline-model-config.h"
#include "sherpa-onnx/csrc/offline-recognizer-impl.h"
#include "sherpa-onnx/csrc/offline-recognizer.h"
#include "sherpa-onnx/csrc/offline-whisper-decoder.h"
#include "sherpa-onnx/csrc/offline-whisper-greedy-search-decoder.h"
#include "sherpa-onnx/csrc/offline-whisper-model.h"
#include "sherpa-onnx/csrc/symbol-table.h"
#include "sherpa-onnx/csrc/transpose.h"
namespace sherpa_onnx {
static OfflineRecognitionResult Convert(const OfflineWhisperDecoderResult &src,
const SymbolTable &sym_table) {
OfflineRecognitionResult r;
r.tokens.reserve(src.tokens.size());
for (auto i : src.tokens) {
if (!sym_table.contains(i)) {
continue;
}
const auto &s = sym_table[i];
r.text += s;
r.tokens.push_back(s);
}
return r;
}
class OfflineRecognizerWhisperImpl : public OfflineRecognizerImpl {
public:
explicit OfflineRecognizerWhisperImpl(const OfflineRecognizerConfig &config)
: config_(config),
symbol_table_(config_.model_config.tokens),
model_(std::make_unique<OfflineWhisperModel>(config.model_config)) {
// tokens.txt from whisper is base64 encoded, so we need to decode it
symbol_table_.ApplyBase64Decode();
if (config.decoding_method == "greedy_search") {
decoder_ =
std::make_unique<OfflineWhisperGreedySearchDecoder>(model_.get());
} else {
SHERPA_ONNX_LOGE(
"Only greedy_search is supported at present for whisper. Given %s",
config.decoding_method.c_str());
exit(-1);
}
}
std::unique_ptr<OfflineStream> CreateStream() const override {
return std::make_unique<OfflineStream>(WhisperTag{});
}
void DecodeStreams(OfflineStream **ss, int32_t n) const override {
// batch decoding is not implemented yet
for (int32_t i = 0; i != n; ++i) {
DecodeStream(ss[i]);
}
}
private:
void DecodeStream(OfflineStream *s) const {
int32_t max_num_frames = 3000;
auto memory_info =
Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
int32_t feat_dim = s->FeatureDim();
std::vector<float> f = s->GetFrames();
int32_t num_frames = f.size() / feat_dim;
if (num_frames > max_num_frames) {
SHERPA_ONNX_LOGE("Only waves less than 30 seconds are supported.");
exit(-1);
}
NormalizeFeatures(f.data(), num_frames, feat_dim);
std::array<int64_t, 3> shape{1, max_num_frames, feat_dim};
Ort::Value mel = Ort::Value::CreateTensor<float>(
model_->Allocator(), shape.data(), shape.size());
float *p_mel = mel.GetTensorMutableData<float>();
std::copy(f.begin(), f.end(), p_mel);
memset(p_mel + f.size(), 0,
(max_num_frames - num_frames) * feat_dim * sizeof(float));
mel = Transpose12(model_->Allocator(), &mel);
auto cross_kv = model_->ForwardEncoder(std::move(mel));
auto results =
decoder_->Decode(std::move(cross_kv.first), std::move(cross_kv.second));
auto r = Convert(results[0], symbol_table_);
s->SetResult(r);
}
private:
static void NormalizeFeatures(float *features, int32_t num_frames,
int32_t feat_dim) {
// log_spec = torch.clamp(features, min=1e-10).log10()
// log_spec = torch.maximum(log_spec, log_spec.max() - 8.0)
// mel = (log_spec + 4.0) / 4.0
int32_t n = num_frames * feat_dim;
float max_v = -1e20;
for (int32_t i = 0; i != n; ++i) {
float f = features[i];
f = std::max<float>(f, 1e-10);
f = std::log10(f);
max_v = std::max(f, max_v);
features[i] = f;
}
max_v -= 8;
for (int32_t i = 0; i != n; ++i) {
float f = features[i];
f = std::max(f, max_v);
f = (f + 4) / 4;
features[i] = f;
}
}
private:
OfflineRecognizerConfig config_;
SymbolTable symbol_table_;
std::unique_ptr<OfflineWhisperModel> model_;
std::unique_ptr<OfflineWhisperDecoder> decoder_;
};
} // namespace sherpa_onnx
#endif // SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_WHISPER_IMPL_H_