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

Fix the last character not being recognized for streaming paraformer models. (#799)

Browse Source
This commit is contained in:
Fangjun Kuang
2024-04-22 15:10:39 +08:00
committed by GitHub
parent 9a68b92ce6
commit 494cb5c733
2 changed files with 27 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
sherpa-onnx/csrc/sherpa-onnx-alsa.cc
View File

@@ -113,6 +113,20 @@ as the device_name.
bool is_endpoint = recognizer.IsEndpoint(stream.get());
if (is_endpoint && !config.model_config.paraformer.encoder.empty()) {
// For streaming paraformer models, since it has a large right chunk size
// we need to pad it on endpointing so that the last character
// can be recognized
std::vector<float> tail_paddings(
static_cast<int>(1.0 * expected_sample_rate));
stream->AcceptWaveform(expected_sample_rate, tail_paddings.data(),
tail_paddings.size());
while (recognizer.IsReady(stream.get())) {
recognizer.DecodeStream(stream.get());
}
text = recognizer.GetResult(stream.get()).text;
}
if (!text.empty() && last_text != text) {
last_text = text;

13
sherpa-onnx/csrc/sherpa-onnx-microphone.cc
View File

@@ -157,6 +157,19 @@ for a list of pre-trained models to download.
auto text = recognizer.GetResult(s.get()).text;
bool is_endpoint = recognizer.IsEndpoint(s.get());
if (is_endpoint && !config.model_config.paraformer.encoder.empty()) {
// For streaming paraformer models, since it has a large right chunk size
// we need to pad it on endpointing so that the last character
// can be recognized
std::vector<float> tail_paddings(static_cast<int>(1.0 * mic_sample_rate));
s->AcceptWaveform(mic_sample_rate, tail_paddings.data(),
tail_paddings.size());
while (recognizer.IsReady(s.get())) {
recognizer.DecodeStream(s.get());
}
text = recognizer.GetResult(s.get()).text;
}
if (!text.empty() && last_text != text) {
last_text = text;