EngineX-Iluvatar/enginex_bi_series-sherpa-onnx
Archived
16
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

Support spoken language identification with whisper (#694)

Browse Source
This commit is contained in:
Fangjun Kuang
2024-03-24 22:57:00 +08:00
committed by GitHub
parent 3cdad9b5d1
commit 0d258dd150
36 changed files with 1173 additions and 200 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
sherpa-onnx/csrc/offline-recognizer-whisper-impl.h
View File

@@ -114,7 +114,7 @@ class OfflineRecognizerWhisperImpl : public OfflineRecognizerImpl {
num_frames = max_num_frames - 50;
}
NormalizeFeatures(f.data(), num_frames, feat_dim);
model_->NormalizeFeatures(f.data(), num_frames, feat_dim);
// note that 1000 is an experience-value.
// You can replace 1000 by other values, say, 100.
@@ -162,38 +162,6 @@ class OfflineRecognizerWhisperImpl : public OfflineRecognizerImpl {
}
}
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_;