Add C++ support for MatchaTTS models not from icefall. (#1834)

.github/scripts/test-offline-tts.sh

@@ -43,6 +43,28 @@ for sid in $(seq 0 10); do
 done
 rm -rf kokoro-en-v0_19
 
+log "------------------------------------------------------------"
+log "matcha-tts-fa_en-male"
+log "------------------------------------------------------------"
+curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/tts-models/matcha-tts-fa_en-male.tar.bz2
+tar xvf matcha-tts-fa_en-male.tar.bz2
+rm matcha-tts-fa_en-male.tar.bz2
+
+curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/vocoder-models/hifigan_v2.onnx
+
+$EXE \
+  --matcha-acoustic-model=./matcha-tts-fa_en-male/model.onnx \
+  --matcha-vocoder=./hifigan_v2.onnx \
+  --matcha-tokens=./matcha-tts-fa_en-male/tokens.txt \
+  --matcha-data-dir=./matcha-tts-fa_en-male/espeak-ng-data \
+  --output-filename=./tts/test-matcha-fa-en-male.wav \
+  --num-threads=2 \
+  "How are you doing today?  این یک نمونه ی تست فارسی است. This is a test."
+
+rm -rf matcha-tts-fa_en-male
+rm hifigan_v2.onnx
+ls -lh tts/*.wav
+
 log "------------------------------------------------------------"
 log "matcha-icefall-en_US-ljspeech"
 log "------------------------------------------------------------"
@@ -64,6 +86,7 @@ $EXE \
 
 rm hifigan_v2.onnx
 rm -rf matcha-icefall-en_US-ljspeech
+ls -lh tts/*.wav
 
 log "------------------------------------------------------------"
 log "matcha-icefall-zh-baker"

scripts/apk/generate-tts-apk-script.py

@@ -397,18 +397,28 @@ def get_matcha_models() -> List[TtsModel]:
         m.dict_dir = m.model_dir + "/dict"
         m.vocoder = "hifigan_v2.onnx"
 
-    english_models = [
+    english_persian_models = [
         TtsModel(
             model_dir="matcha-icefall-en_US-ljspeech",
             acoustic_model_name="model-steps-3.onnx",
             lang="en",
-        )
+        ),
+        TtsModel(
+            model_dir="matcha-tts-fa_en-male",
+            acoustic_model_name="model.onnx",
+            lang="fa",
+        ),
+        TtsModel(
+            model_dir="matcha-tts-fa_en-female",
+            acoustic_model_name="model.onnx",
+            lang="fa",
+        ),
     ]
-    for m in english_models:
+    for m in english_persian_models:
         m.data_dir = f"{m.model_dir}/espeak-ng-data"
         m.vocoder = "hifigan_v2.onnx"
 
-    return chinese_models + english_models
+    return chinese_models + english_persian_models
 
 
 def get_kokoro_models() -> List[TtsModel]:

sherpa-onnx/csrc/offline-tts-matcha-impl.h

@@ -214,7 +214,7 @@ class OfflineTtsMatchaImpl : public OfflineTtsImpl {
     }
 
     std::vector<TokenIDs> token_ids =
-        frontend_->ConvertTextToTokenIds(text, "en-US");
+        frontend_->ConvertTextToTokenIds(text, meta_data.voice);
 
     if (token_ids.empty() ||
         (token_ids.size() == 1 && token_ids[0].tokens.empty())) {

sherpa-onnx/csrc/offline-tts-matcha-model-meta-data.h

@@ -21,6 +21,8 @@ struct OfflineTtsMatchaModelMetaData {
   int32_t has_espeak = 0;
   int32_t use_eos_bos = 0;
   int32_t pad_id = 0;
+
+  std::string voice;
 };
 
 }  // namespace sherpa_onnx

sherpa-onnx/csrc/offline-tts-matcha-model.cc

@@ -83,15 +83,32 @@ class OfflineTtsMatchaModel::Impl {
     Ort::Value sid_tensor =
         Ort::Value::CreateTensor(memory_info, &sid, 1, &scale_shape, 1);
 
+    std::array<float, 2> scales = {noise_scale, length_scale};
+    int64_t scales_shape = 2;
+
+    Ort::Value scales_tensor = Ort::Value::CreateTensor(
+        memory_info, scales.data(), scales.size(), &scales_shape, 1);
+
     std::vector<Ort::Value> inputs;
     inputs.reserve(5);
     inputs.push_back(std::move(x));
     inputs.push_back(std::move(x_length));
-    inputs.push_back(std::move(noise_scale_tensor));
-    inputs.push_back(std::move(length_scale_tensor));
+    if (input_names_[2] == "scales") {
+      // for models from
+      // https://github.com/shivammehta25/Matcha-TTS
+      inputs.push_back(std::move(scales_tensor));
+    } else {
+      // for models from icefall
+      inputs.push_back(std::move(noise_scale_tensor));
+      inputs.push_back(std::move(length_scale_tensor));
+    }
 
     if (input_names_.size() == 5 && input_names_.back() == "sid") {
+      // for models from icefall
       inputs.push_back(std::move(sid_tensor));
+
+      // Note that we have not supported multi-speaker tts models from
+      // https://github.com/shivammehta25/Matcha-TTS
     }
 
     auto out =
@@ -145,6 +162,8 @@ class OfflineTtsMatchaModel::Impl {
     SHERPA_ONNX_READ_META_DATA(meta_data_.has_espeak, "has_espeak");
     SHERPA_ONNX_READ_META_DATA(meta_data_.use_eos_bos, "use_eos_bos");
     SHERPA_ONNX_READ_META_DATA(meta_data_.pad_id, "pad_id");
+    SHERPA_ONNX_READ_META_DATA_STR_WITH_DEFAULT(meta_data_.voice, "voice",
+                                                "en-us");
   }
 
  private: