Add C API for streaming HLG decoding (#734)

This commit is contained in:
Fangjun Kuang
2024-04-05 10:31:20 +08:00
committed by GitHub
parent db67e00c77
commit dbff2eaadb
39 changed files with 839 additions and 8 deletions

View File

@@ -43,6 +43,10 @@ function freeConfig(config, Module) {
freeConfig(config.lm, Module)
}
if ('ctcFstDecoder' in config) {
freeConfig(config.ctcFstDecoder, Module)
}
Module._free(config.ptr);
}
@@ -193,11 +197,26 @@ function initSherpaOnnxFeatureConfig(config, Module) {
return {ptr: ptr, len: len};
}
function initSherpaOnnxOnlineCtcFstDecoderConfig(config, Module) {
const len = 2 * 4;
const ptr = Module._malloc(len);
const graphLen = Module.lengthBytesUTF8(config.graph) + 1;
const buffer = Module._malloc(graphLen);
Module.stringToUTF8(config.graph, buffer, graphLen);
Module.setValue(ptr, buffer, 'i8*');
Module.setValue(ptr + 4, config.maxActive, 'i32');
return {ptr: ptr, len: len, buffer: buffer};
}
function initSherpaOnnxOnlineRecognizerConfig(config, Module) {
const feat = initSherpaOnnxFeatureConfig(config.featConfig, Module);
const model = initSherpaOnnxOnlineModelConfig(config.modelConfig, Module);
const ctcFstDecoder = initSherpaOnnxOnlineCtcFstDecoderConfig(
config.ctcFstDecoderConfig, Module)
const len = feat.len + model.len + 8 * 4;
const len = feat.len + model.len + 8 * 4 + ctcFstDecoder.len;
const ptr = Module._malloc(len);
let offset = 0;
@@ -243,8 +262,11 @@ function initSherpaOnnxOnlineRecognizerConfig(config, Module) {
Module.setValue(ptr + offset, config.hotwordsScore, 'float');
offset += 4;
Module._CopyHeap(ctcFstDecoder.ptr, ctcFstDecoder.len, ptr + offset);
return {
buffer: buffer, ptr: ptr, len: len, feat: feat, model: model
buffer: buffer, ptr: ptr, len: len, feat: feat, model: model,
ctcFstDecoder: ctcFstDecoder
}
}
@@ -313,6 +335,10 @@ function createOnlineRecognizer(Module, myConfig) {
rule3MinUtteranceLength: 20,
hotwordsFile: '',
hotwordsScore: 1.5,
ctcFstDecoderConfig: {
graph: '',
maxActive: 3000,
}
};
if (myConfig) {
recognizerConfig = myConfig;

View File

@@ -22,9 +22,11 @@ static_assert(sizeof(SherpaOnnxOnlineModelConfig) ==
sizeof(SherpaOnnxOnlineZipformer2CtcModelConfig) + 5 * 4,
"");
static_assert(sizeof(SherpaOnnxFeatureConfig) == 2 * 4, "");
static_assert(sizeof(SherpaOnnxOnlineCtcFstDecoderConfig) == 2 * 4, "");
static_assert(sizeof(SherpaOnnxOnlineRecognizerConfig) ==
sizeof(SherpaOnnxFeatureConfig) +
sizeof(SherpaOnnxOnlineModelConfig) + 8 * 4,
sizeof(SherpaOnnxOnlineModelConfig) + 8 * 4 +
sizeof(SherpaOnnxOnlineCtcFstDecoderConfig),
"");
void MyPrint(SherpaOnnxOnlineRecognizerConfig *config) {
@@ -67,6 +69,11 @@ void MyPrint(SherpaOnnxOnlineRecognizerConfig *config) {
config->rule3_min_utterance_length);
fprintf(stdout, "hotwords_file: %s\n", config->hotwords_file);
fprintf(stdout, "hotwords_score: %.2f\n", config->hotwords_score);
fprintf(stdout, "----------ctc fst decoder config----------\n");
fprintf(stdout, "graph: %s\n", config->ctc_fst_decoder_config.graph);
fprintf(stdout, "max_active: %d\n",
config->ctc_fst_decoder_config.max_active);
}
void CopyHeap(const char *src, int32_t num_bytes, char *dst) {