Add CXX API for MatchaTTS models (#1676)

This commit is contained in:
Fangjun Kuang
2025-01-03 14:16:36 +08:00
committed by GitHub
parent 9aa4897a9e
commit 648903834b
12 changed files with 403 additions and 8 deletions

View File

@@ -1114,7 +1114,7 @@ static sherpa_onnx::OfflineTtsConfig GetOfflineTtsConfig(
return tts_config;
}
SherpaOnnxOfflineTts *SherpaOnnxCreateOfflineTts(
const SherpaOnnxOfflineTts *SherpaOnnxCreateOfflineTts(
const SherpaOnnxOfflineTtsConfig *config) {
auto tts_config = GetOfflineTtsConfig(config);
@@ -1130,7 +1130,9 @@ SherpaOnnxOfflineTts *SherpaOnnxCreateOfflineTts(
return tts;
}
void SherpaOnnxDestroyOfflineTts(SherpaOnnxOfflineTts *tts) { delete tts; }
void SherpaOnnxDestroyOfflineTts(const SherpaOnnxOfflineTts *tts) {
delete tts;
}
int32_t SherpaOnnxOfflineTtsSampleRate(const SherpaOnnxOfflineTts *tts) {
return tts->impl->SampleRate();

View File

@@ -950,11 +950,12 @@ SHERPA_ONNX_API typedef struct SherpaOnnxOfflineTts SherpaOnnxOfflineTts;
// Create an instance of offline TTS. The user has to use DestroyOfflineTts()
// to free the returned pointer to avoid memory leak.
SHERPA_ONNX_API SherpaOnnxOfflineTts *SherpaOnnxCreateOfflineTts(
SHERPA_ONNX_API const SherpaOnnxOfflineTts *SherpaOnnxCreateOfflineTts(
const SherpaOnnxOfflineTtsConfig *config);
// Free the pointer returned by SherpaOnnxCreateOfflineTts()
SHERPA_ONNX_API void SherpaOnnxDestroyOfflineTts(SherpaOnnxOfflineTts *tts);
SHERPA_ONNX_API void SherpaOnnxDestroyOfflineTts(
const SherpaOnnxOfflineTts *tts);
// Return the sample rate of the current TTS object
SHERPA_ONNX_API int32_t
@@ -984,7 +985,6 @@ SHERPA_ONNX_API
const SherpaOnnxGeneratedAudio *
SherpaOnnxOfflineTtsGenerateWithProgressCallback(
const SherpaOnnxOfflineTts *tts, const char *text, int32_t sid, float speed,
SherpaOnnxGeneratedAudioProgressCallback callback);
SHERPA_ONNX_API

View File

@@ -24,6 +24,11 @@ Wave ReadWave(const std::string &filename) {
return ans;
}
bool WriteWave(const std::string &filename, const Wave &wave) {
return SherpaOnnxWriteWave(wave.samples.data(), wave.samples.size(),
wave.sample_rate, filename.c_str());
}
OnlineStream::OnlineStream(const SherpaOnnxOnlineStream *p)
: MoveOnly<OnlineStream, SherpaOnnxOnlineStream>(p) {}
@@ -311,4 +316,73 @@ OfflineRecognizerResult OfflineRecognizer::GetResult(
return ans;
}
OfflineTts OfflineTts::Create(const OfflineTtsConfig &config) {
struct SherpaOnnxOfflineTtsConfig c;
memset(&c, 0, sizeof(c));
c.model.vits.model = config.model.vits.model.c_str();
c.model.vits.lexicon = config.model.vits.lexicon.c_str();
c.model.vits.tokens = config.model.vits.tokens.c_str();
c.model.vits.data_dir = config.model.vits.data_dir.c_str();
c.model.vits.noise_scale = config.model.vits.noise_scale;
c.model.vits.noise_scale_w = config.model.vits.noise_scale_w;
c.model.vits.length_scale = config.model.vits.length_scale;
c.model.vits.dict_dir = config.model.vits.dict_dir.c_str();
c.model.matcha.acoustic_model = config.model.matcha.acoustic_model.c_str();
c.model.matcha.vocoder = config.model.matcha.vocoder.c_str();
c.model.matcha.lexicon = config.model.matcha.lexicon.c_str();
c.model.matcha.tokens = config.model.matcha.tokens.c_str();
c.model.matcha.data_dir = config.model.matcha.data_dir.c_str();
c.model.matcha.noise_scale = config.model.matcha.noise_scale;
c.model.matcha.length_scale = config.model.matcha.length_scale;
c.model.matcha.dict_dir = config.model.matcha.dict_dir.c_str();
c.model.num_threads = config.model.num_threads;
c.model.debug = config.model.debug;
c.model.provider = config.model.provider.c_str();
c.rule_fsts = config.rule_fsts.c_str();
c.max_num_sentences = config.max_num_sentences;
c.rule_fars = config.rule_fars.c_str();
auto p = SherpaOnnxCreateOfflineTts(&c);
return OfflineTts(p);
}
OfflineTts::OfflineTts(const SherpaOnnxOfflineTts *p)
: MoveOnly<OfflineTts, SherpaOnnxOfflineTts>(p) {}
void OfflineTts::Destroy(const SherpaOnnxOfflineTts *p) const {
SherpaOnnxDestroyOfflineTts(p);
}
int32_t OfflineTts::SampleRate() const {
return SherpaOnnxOfflineTtsSampleRate(p_);
}
int32_t OfflineTts::NumSpeakers() const {
return SherpaOnnxOfflineTtsNumSpeakers(p_);
}
GeneratedAudio OfflineTts::Generate(const std::string &text,
int32_t sid /*= 0*/, float speed /*= 1.0*/,
OfflineTtsCallback callback /*= nullptr*/,
void *arg /*= nullptr*/) const {
const SherpaOnnxGeneratedAudio *audio;
if (!callback) {
audio = SherpaOnnxOfflineTtsGenerate(p_, text.c_str(), sid, speed);
} else {
audio = SherpaOnnxOfflineTtsGenerateWithProgressCallbackWithArg(
p_, text.c_str(), sid, speed, callback, arg);
}
GeneratedAudio ans;
ans.samples = std::vector<float>{audio->samples, audio->samples + audio->n};
ans.sample_rate = audio->sample_rate;
SherpaOnnxDestroyOfflineTtsGeneratedAudio(audio);
return ans;
}
} // namespace sherpa_onnx::cxx

View File

@@ -97,6 +97,10 @@ struct Wave {
SHERPA_ONNX_API Wave ReadWave(const std::string &filename);
// Return true on success;
// Return false on failure
SHERPA_ONNX_API bool WriteWave(const std::string &filename, const Wave &wave);
template <typename Derived, typename T>
class SHERPA_ONNX_API MoveOnly {
public:
@@ -307,6 +311,91 @@ class SHERPA_ONNX_API OfflineRecognizer
explicit OfflineRecognizer(const SherpaOnnxOfflineRecognizer *p);
};
// ============================================================================
// Non-streaming TTS
// ============================================================================
struct OfflineTtsVitsModelConfig {
std::string model;
std::string lexicon;
std::string tokens;
std::string data_dir;
std::string dict_dir;
float noise_scale = 0.667;
float noise_scale_w = 0.8;
float length_scale = 1.0; // < 1, faster in speed; > 1, slower in speed
};
struct OfflineTtsMatchaModelConfig {
std::string acoustic_model;
std::string vocoder;
std::string lexicon;
std::string tokens;
std::string data_dir;
std::string dict_dir;
float noise_scale = 0.667;
float length_scale = 1.0; // < 1, faster in speed; > 1, slower in speed
};
struct OfflineTtsModelConfig {
OfflineTtsVitsModelConfig vits;
OfflineTtsMatchaModelConfig matcha;
int32_t num_threads = 1;
bool debug = false;
std::string provider = "cpu";
};
struct OfflineTtsConfig {
OfflineTtsModelConfig model;
std::string rule_fsts;
std::string rule_fars;
int32_t max_num_sentences = 1;
};
struct GeneratedAudio {
std::vector<float> samples; // in the range [-1, 1]
int32_t sample_rate;
};
// Return 1 to continue generating
// Return 0 to stop generating
using OfflineTtsCallback = int32_t (*)(const float *samples,
int32_t num_samples, float progress,
void *arg);
class SHERPA_ONNX_API OfflineTts
: public MoveOnly<OfflineTts, SherpaOnnxOfflineTts> {
public:
static OfflineTts Create(const OfflineTtsConfig &config);
void Destroy(const SherpaOnnxOfflineTts *p) const;
// Return the sample rate of the generated audio
int32_t SampleRate() const;
// Number of supported speakers.
// If it supports only a single speaker, then it return 0 or 1.
int32_t NumSpeakers() const;
// @param text A string containing words separated by spaces
// @param sid Speaker ID. Used only for multi-speaker models, e.g., models
// trained using the VCTK dataset. It is not used for
// single-speaker models, e.g., models trained using the ljspeech
// dataset.
// @param speed The speed for the generated speech. E.g., 2 means 2x faster.
// @param callback If not NULL, it is called whenever config.max_num_sentences
// sentences have been processed. The callback is called in
// the current thread.
GeneratedAudio Generate(const std::string &text, int32_t sid = 0,
float speed = 1.0,
OfflineTtsCallback callback = nullptr,
void *arg = nullptr) const;
private:
explicit OfflineTts(const SherpaOnnxOfflineTts *p);
};
} // namespace sherpa_onnx::cxx
#endif // SHERPA_ONNX_C_API_CXX_API_H_