11// MIT License
22
3- // Copyright (c) 2023 Miguel Ángel González Santamarta
3+ // Copyright (c) 2024 Miguel Ángel González Santamarta
44
55// Permission is hereby granted, free of charge, to any person obtaining a copy
66// of this software and associated documentation files (the "Software"), to deal
2020// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
2121// SOFTWARE.
2222
23+ #include < algorithm>
2324#include < limits>
2425#include < memory>
2526#include < string>
@@ -31,32 +32,39 @@ using namespace silero_vad;
3132
3233VadIterator::VadIterator (const std::string &model_path, int sample_rate,
3334 int frame_size_ms, float threshold, int min_silence_ms,
34- int speech_pad_ms, int min_speech_ms,
35- float max_speech_s)
35+ int speech_pad_ms)
3636 : env(ORT_LOGGING_LEVEL_WARNING, " VadIterator"
3737 sample_rate(sample_rate), sr_per_ms(sample_rate / 1000 ),
3838 window_size_samples(frame_size_ms * sr_per_ms),
39- min_speech_samples(sr_per_ms * min_speech_ms),
4039 speech_pad_samples(sr_per_ms * speech_pad_ms),
41- max_speech_samples(sample_rate * max_speech_s - window_size_samples -
42- 2 * speech_pad_samples),
4340 min_silence_samples(sr_per_ms * min_silence_ms),
44- min_silence_samples_at_max_speech(sr_per_ms * 98 ),
45- state(2 * 1 * 128 , 0 .0f ), sr(1 , sample_rate), context( 64 , 0 . 0f ) {
41+ context_size(sample_rate == 16000 ? 64 : 32 ), context(context_size, 0 . 0f ),
42+ state(2 * 1 * 128 , 0 .0f ), sr(1 , sample_rate) {
4643
47- // this->input.resize(window_size_samples);
4844 this ->input_node_dims [0 ] = 1 ;
4945 this ->input_node_dims [1 ] = window_size_samples;
50- this ->init_onnx_model (model_path);
46+
47+ try {
48+ this ->init_onnx_model (model_path);
49+ } catch (const std::exception &e) {
50+ throw std::runtime_error (" Failed to initialize ONNX model: "
51+ std::string (e.what ()));
52+ }
5153}
5254
5355void VadIterator::init_onnx_model (const std::string &model_path) {
5456 this ->session_options .SetIntraOpNumThreads (1 );
5557 this ->session_options .SetInterOpNumThreads (1 );
5658 this ->session_options .SetGraphOptimizationLevel (
5759 GraphOptimizationLevel::ORT_ENABLE_ALL);
58- this ->session = std::make_shared<Ort::Session>(this ->env , model_path.c_str (),
59- this ->session_options );
60+
61+ try {
62+ this ->session = std::make_shared<Ort::Session>(
63+ this ->env , model_path.c_str (), this ->session_options );
64+ } catch (const std::exception &e) {
65+ throw std::runtime_error (" Failed to create ONNX session: "
66+ std::string (e.what ()));
67+ }
6068}
6169
6270void VadIterator::reset_states () {
@@ -68,16 +76,13 @@ void VadIterator::reset_states() {
6876}
6977
7078Timestamp VadIterator::predict (const std::vector<float > &data) {
71- // Create input tensors
79+ // Pre-fill input with context
7280 this ->input .clear ();
73- for (auto ele : this ->context ) {
74- this ->input .push_back (ele);
75- }
76-
77- for (auto ele : data) {
78- this ->input .push_back (ele);
79- }
81+ this ->input .reserve (context.size () + data.size ());
82+ this ->input .insert (input.end (), context.begin (), context.end ());
83+ this ->input .insert (input.end (), data.begin (), data.end ());
8084
85+ // Create input tensors
8186 Ort::Value input_tensor = Ort::Value::CreateTensor<float >(
8287 this ->memory_info , this ->input .data (), this ->input .size (),
8388 this ->input_node_dims , 2 );
@@ -95,20 +100,23 @@ Timestamp VadIterator::predict(const std::vector<float> &data) {
95100 this ->ort_inputs .emplace_back (std::move (sr_tensor));
96101
97102 // Run inference
98- this ->ort_outputs = this ->session ->Run (
99- Ort::RunOptions{nullptr }, this ->input_node_names .data (),
100- this ->ort_inputs .data (), this ->ort_inputs .size (),
101- this ->output_node_names .data (), this ->output_node_names .size ());
103+ try {
104+ this ->ort_outputs = session->Run (
105+ Ort::RunOptions{nullptr }, this ->input_node_names .data (),
106+ this ->ort_inputs .data (), this ->ort_inputs .size (),
107+ this ->output_node_names .data (), this ->output_node_names .size ());
108+ } catch (const std::exception &e) {
109+ throw std::runtime_error (" ONNX inference failed: " std::string (e.what ()));
110+ }
102111
103112 // Process output
104113 float speech_prob = this ->ort_outputs [0 ].GetTensorMutableData <float >()[0 ];
105114 float *updated_state = this ->ort_outputs [1 ].GetTensorMutableData <float >();
106115 std::copy (updated_state, updated_state + this ->state .size (),
107116 this ->state .begin ());
108117
109- for (int i = 64 ; i > 0 ; i--) {
110- this ->context .push_back (data.at (data.size () - i));
111- }
118+ // Update context with the last 64 samples of data
119+ this ->context .assign (data.end () - context_size, data.end ());
112120
113121 // Handle result
114122 this ->current_sample += this ->window_size_samples ;
@@ -119,10 +127,10 @@ Timestamp VadIterator::predict(const std::vector<float> &data) {
119127 }
120128
121129 if (!this ->triggered ) {
130+ int start_timestwamp = this ->current_sample - this ->speech_pad_samples -
131+ this ->window_size_samples ;
122132 this ->triggered = true ;
123- return Timestamp (this ->current_sample - this ->speech_pad_samples -
124- this ->window_size_samples ,
125- -1 , speech_prob);
133+ return Timestamp (start_timestwamp, -1 , speech_prob);
126134 }
127135
128136 } else if (speech_prob < this ->threshold - 0.15 && this ->triggered ) {
@@ -131,12 +139,11 @@ Timestamp VadIterator::predict(const std::vector<float> &data) {
131139 }
132140
133141 if (this ->current_sample - this ->temp_end >= this ->min_silence_samples ) {
134- this ->temp_end = 0 ;
142+ int end_timestamp =
143+ this ->temp_end + this ->speech_pad_samples - this ->window_size_samples ;
135144 this ->triggered = false ;
136- return Timestamp (-1 ,
137- this ->temp_end + this ->speech_pad_samples -
138- this ->window_size_samples ,
139- speech_prob);
145+ this ->temp_end = 0 ;
146+ return Timestamp (-1 , end_timestamp, speech_prob);
140147 }
141148 }
142149
0 commit comments