jamosdev
/
rhubarb-lip-sync
mirror of https://github.com/DanielSWolf/rhubarb-lip-sync.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Using biased language model to handle dialog more forgivingly 

Using a fixed 0.1-0.9 ratio between default and dialog language model
This commit is contained in:
Daniel Wolf 2016-10-21 21:41:50 +02:00
parent 9cfe577612
commit c19ad1c8d0
1 changed files with 46 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

62
src/phoneRecognition.cpp
View File

@ -45,6 +45,7 @@ using std::chrono::duration;
using boost::optional; using boost::optional;
using std::u32string; using std::u32string;
using std::chrono::duration_cast; using std::chrono::duration_cast;
using std::array;
constexpr int sphinxSampleRate = 16000; constexpr int sphinxSampleRate = 16000;
@ -237,6 +238,48 @@ void addMissingDictionaryWords(const vector<string>& words, ps_decoder_t& decode
} }
} }
lambda_unique_ptr<ngram_model_t> createDefaultLanguageModel(ps_decoder_t& decoder) {
path modelPath = getSphinxModelDirectory() / "en-us.lm.bin";
lambda_unique_ptr<ngram_model_t> result(
ngram_model_read(decoder.config, modelPath.string().c_str(), NGRAM_AUTO, decoder.lmath),
[](ngram_model_t* lm) { ngram_model_free(lm); });
if (!result) {
throw runtime_error(fmt::format("Error reading language model from {}.", modelPath));
}
return std::move(result);
}
lambda_unique_ptr<ngram_model_t> createDialogLanguageModel(ps_decoder_t& decoder, const u32string& dialog) {
// Split dialog into normalized words
vector<string> words = tokenizeText(dialog, [&](const string& word) { return dictionaryContains(*decoder.dict, word); });
// Add dialog-specific words to the dictionary
addMissingDictionaryWords(words, decoder);
// Create dialog-specific language model
words.insert(words.begin(), "<s>");
words.push_back("</s>");
return createLanguageModel(words, decoder);
}
lambda_unique_ptr<ngram_model_t> createBiasedLanguageModel(ps_decoder_t& decoder, const u32string& dialog) {
auto defaultLanguageModel = createDefaultLanguageModel(decoder);
auto dialogLanguageModel = createDialogLanguageModel(decoder, dialog);
constexpr int modelCount = 2;
array<ngram_model_t*, modelCount> languageModels{ defaultLanguageModel.get(), dialogLanguageModel.get() };
array<char*, modelCount> modelNames{ "defaultLM", "dialogLM" };
array<float, modelCount> modelWeights{ 0.1f, 0.9f };
lambda_unique_ptr<ngram_model_t> result(
ngram_model_set_init(nullptr, languageModels.data(), modelNames.data(), modelWeights.data(), modelCount),
[](ngram_model_t* lm) { ngram_model_free(lm); });
if (!result) {
throw runtime_error("Error creating biased language model.");
}
return std::move(result);
}
lambda_unique_ptr<ps_decoder_t> createDecoder(optional<u32string> dialog) { lambda_unique_ptr<ps_decoder_t> createDecoder(optional<u32string> dialog) {
lambda_unique_ptr<cmd_ln_t> config( lambda_unique_ptr<cmd_ln_t> config(
cmd_ln_init( cmd_ln_init(
@ -261,22 +304,9 @@ lambda_unique_ptr<ps_decoder_t> createDecoder(optional<u32string> dialog) {
if (!decoder) throw runtime_error("Error creating speech decoder."); if (!decoder) throw runtime_error("Error creating speech decoder.");
// Set language model // Set language model
lambda_unique_ptr<ngram_model_t> languageModel; lambda_unique_ptr<ngram_model_t> languageModel(dialog
if (dialog) { ? createBiasedLanguageModel(*decoder, *dialog)
// Create dialog-specific language model : createDefaultLanguageModel(*decoder));
vector<string> words = tokenizeText(*dialog, [&](const string& word) { return dictionaryContains(*decoder->dict, word); });
words.insert(words.begin(), "<s>");
words.push_back("</s>");
languageModel = createLanguageModel(words, *decoder);
// Add any dialog-specific words to the dictionary
addMissingDictionaryWords(words, *decoder);
} else {
path modelPath = getSphinxModelDirectory() / "en-us.lm.bin";
languageModel = lambda_unique_ptr<ngram_model_t>(
ngram_model_read(decoder->config, modelPath.string().c_str(), NGRAM_AUTO, decoder->lmath),
[](ngram_model_t* lm) { ngram_model_free(lm); });
}
ps_set_lm(decoder.get(), "lm", languageModel.get()); ps_set_lm(decoder.get(), "lm", languageModel.get());
ps_set_search(decoder.get(), "lm"); ps_set_search(decoder.get(), "lm");