Support scaling the duration of a pause in TTS. (#1820)

@@ -116,7 +116,7 @@ int32_t main() {
   keywords_spotter_config.keywords_buf = keywords_buf;
   keywords_spotter_config.keywords_buf_size = keywords_buf_size;
 
-  SherpaOnnxKeywordSpotter *keywords_spotter =
+  const SherpaOnnxKeywordSpotter *keywords_spotter =
       SherpaOnnxCreateKeywordSpotter(&keywords_spotter_config);
 
   free((void *)tokens_buf);
@@ -130,7 +130,7 @@ int32_t main() {
     return -1;
   }
 
-  SherpaOnnxOnlineStream *stream =
+  const SherpaOnnxOnlineStream *stream =
       SherpaOnnxCreateKeywordStream(keywords_spotter);
 
   const SherpaOnnxDisplay *display = SherpaOnnxCreateDisplay(50);

@@ -180,6 +180,9 @@ final class SherpaOnnxOfflineTtsConfig extends Struct {
   external int maxNumSenetences;
 
   external Pointer<Utf8> ruleFars;
+
+  @Float()
+  external double silenceScale;
 }
 
 final class SherpaOnnxGeneratedAudio extends Struct {

@@ -114,17 +114,19 @@ class OfflineTtsConfig {
     this.ruleFsts = '',
     this.maxNumSenetences = 1,
     this.ruleFars = '',
+    this.silenceScale = 0.2,
   });
 
   @override
   String toString() {
-    return 'OfflineTtsConfig(model: $model, ruleFsts: $ruleFsts, maxNumSenetences: $maxNumSenetences, ruleFars: $ruleFars)';
+    return 'OfflineTtsConfig(model: $model, ruleFsts: $ruleFsts, maxNumSenetences: $maxNumSenetences, ruleFars: $ruleFars, silenceScale: $silenceScale)';
   }
 
   final OfflineTtsModelConfig model;
   final String ruleFsts;
   final int maxNumSenetences;
   final String ruleFars;
+  final double silenceScale;
 }
 
 class GeneratedAudio {
@@ -180,6 +182,7 @@ class OfflineTts {
     c.ref.ruleFsts = config.ruleFsts.toNativeUtf8();
     c.ref.maxNumSenetences = config.maxNumSenetences;
     c.ref.ruleFars = config.ruleFars.toNativeUtf8();
+    c.ref.silenceScale = config.silenceScale;
 
     final ptr = SherpaOnnxBindings.createOfflineTts?.call(c) ?? nullptr;
 

@@ -146,6 +146,7 @@ static Napi::External<SherpaOnnxOfflineTts> CreateOfflineTtsWrapper(
   SHERPA_ONNX_ASSIGN_ATTR_STR(rule_fsts, ruleFsts);
   SHERPA_ONNX_ASSIGN_ATTR_INT32(max_num_sentences, maxNumSentences);
   SHERPA_ONNX_ASSIGN_ATTR_STR(rule_fars, ruleFars);
+  SHERPA_ONNX_ASSIGN_ATTR_STR(silence_scale, silenceScale);
 
 #if __OHOS__
   std::unique_ptr<NativeResourceManager,

@@ -52,6 +52,7 @@ export class OfflineTtsConfig {
   public ruleFsts: string = '';
   public ruleFars: string = '';
   public maxNumSentences: number = 1;
+  public silenceScale: number = 0.2;
 }
 
 export class TtsOutput {
@@ -98,4 +99,4 @@ export class OfflineTts {
   generateAsync(input: TtsInput): Promise<TtsOutput> {
     return offlineTtsGenerateAsync(this.handle, input);
   }
-}
\ No newline at end of file
+}

@@ -13,6 +13,7 @@ namespace SherpaOnnx
             RuleFsts = "";
             MaxNumSentences = 1;
             RuleFars = "";
+            SilenceScale = 0.2F;
         }
         public OfflineTtsModelConfig Model;
 
@@ -23,6 +24,7 @@ namespace SherpaOnnx
 
         [MarshalAs(UnmanagedType.LPStr)]
         public string RuleFars;
-    }
 
-}
\ No newline at end of file
+        public float SilenceScale;
+    }
+}

@@ -712,6 +712,7 @@ type OfflineTtsConfig struct {
 	RuleFsts        string
 	RuleFars        string
 	MaxNumSentences int
+	SilenceScale    float32
 }
 
 type GeneratedAudio struct {
@@ -744,6 +745,7 @@ func NewOfflineTts(config *OfflineTtsConfig) *OfflineTts {
 	defer C.free(unsafe.Pointer(c.rule_fars))
 
 	c.max_num_sentences = C.int(config.MaxNumSentences)
+	c.silence_scale = C.float(config.SilenceScale)
 
 	// vits
 	c.model.vits.model = C.CString(config.Model.Vits.Model)

@@ -1135,6 +1135,7 @@ static sherpa_onnx::OfflineTtsConfig GetOfflineTtsConfig(
   tts_config.rule_fsts = SHERPA_ONNX_OR(config->rule_fsts, "");
   tts_config.rule_fars = SHERPA_ONNX_OR(config->rule_fars, "");
   tts_config.max_num_sentences = SHERPA_ONNX_OR(config->max_num_sentences, 1);
+  tts_config.silence_scale = SHERPA_ONNX_OR(config->silence_scale, 0.2);
 
   if (tts_config.model.debug) {
 #if __OHOS__

@@ -944,6 +944,7 @@ SHERPA_ONNX_API typedef struct SherpaOnnxOfflineTtsConfig {
   const char *rule_fsts;
   int32_t max_num_sentences;
   const char *rule_fars;
+  float silence_scale;
 } SherpaOnnxOfflineTtsConfig;
 
 SHERPA_ONNX_API typedef struct SherpaOnnxGeneratedAudio {

@@ -352,6 +352,7 @@ OfflineTts OfflineTts::Create(const OfflineTtsConfig &config) {
 
   c.rule_fsts = config.rule_fsts.c_str();
   c.max_num_sentences = config.max_num_sentences;
+  c.silence_scale = config.silence_scale;
   c.rule_fars = config.rule_fars.c_str();
 
   auto p = SherpaOnnxCreateOfflineTts(&c);

@@ -363,6 +363,7 @@ struct OfflineTtsConfig {
   std::string rule_fsts;
   std::string rule_fars;
   int32_t max_num_sentences = 1;
+  float silence_scale = 0.2;
 };
 
 struct GeneratedAudio {

@@ -420,6 +420,12 @@ class OfflineTtsKokoroImpl : public OfflineTtsImpl {
     GeneratedAudio ans;
     ans.sample_rate = model_->GetMetaData().sample_rate;
     ans.samples = std::vector<float>(p, p + total);
+
+    float silence_scale = config_.silence_scale;
+    if (silence_scale != 1) {
+      ans = ans.ScaleSilence(silence_scale);
+    }
+
     return ans;
   }
 

@@ -398,6 +398,12 @@ class OfflineTtsMatchaImpl : public OfflineTtsImpl {
     GeneratedAudio ans;
     ans.sample_rate = model_->GetMetaData().sample_rate;
     ans.samples = std::vector<float>(p, p + total);
+
+    float silence_scale = config_.silence_scale;
+    if (silence_scale != 1) {
+      ans = ans.ScaleSilence(silence_scale);
+    }
+
     return ans;
   }
 

@@ -485,6 +485,12 @@ class OfflineTtsVitsImpl : public OfflineTtsImpl {
     GeneratedAudio ans;
     ans.sample_rate = model_->GetMetaData().sample_rate;
     ans.samples = std::vector<float>(p, p + total);
+
+    float silence_scale = config_.silence_scale;
+    if (silence_scale != 1) {
+      ans = ans.ScaleSilence(silence_scale);
+    }
+
     return ans;
   }
 

@@ -4,6 +4,7 @@
 
 #include "sherpa-onnx/csrc/offline-tts.h"
 
+#include <cmath>
 #include <string>
 #include <utility>
 
@@ -23,6 +24,72 @@
 
 namespace sherpa_onnx {
 
+struct SilenceInterval {
+  int32_t start;
+  int32_t end;
+};
+
+GeneratedAudio GeneratedAudio::ScaleSilence(float scale) const {
+  if (scale == 1) {
+    return *this;
+  }
+  // if the interval is larger than 0.6 second, then we assume it is a pause
+  int32_t threshold = static_cast<int32_t>(sample_rate * 0.6);
+
+  std::vector<SilenceInterval> intervals;
+  int32_t num_samples = static_cast<int32_t>(samples.size());
+
+  int32_t last = -1;
+  int32_t i;
+  for (i = 0; i != num_samples; ++i) {
+    if (fabs(samples[i]) <= 0.01) {
+      if (last == -1) {
+        last = i;
+      }
+      continue;
+    }
+
+    if (last != -1 && i - last < threshold) {
+      last = -1;
+      continue;
+    }
+
+    if (last != -1) {
+      intervals.push_back({last, i});
+      last = -1;
+    }
+  }
+
+  if (last != -1 && num_samples - last > threshold) {
+    intervals.push_back({last, num_samples});
+  }
+
+  if (intervals.empty()) {
+    return *this;
+  }
+
+  GeneratedAudio ans;
+  ans.sample_rate = sample_rate;
+  ans.samples.reserve(samples.size());
+
+  i = 0;
+  for (const auto &interval : intervals) {
+    ans.samples.insert(ans.samples.end(), samples.begin() + i,
+                       samples.begin() + interval.start);
+    i = interval.end;
+    int32_t n = static_cast<int32_t>((interval.end - interval.start) * scale);
+
+    ans.samples.insert(ans.samples.end(), samples.begin() + interval.start,
+                       samples.begin() + interval.start + n);
+  }
+
+  if (i < num_samples) {
+    ans.samples.insert(ans.samples.end(), samples.begin() + i, samples.end());
+  }
+
+  return ans;
+}
+
 void OfflineTtsConfig::Register(ParseOptions *po) {
   model.Register(po);
 
@@ -44,6 +111,10 @@ void OfflineTtsConfig::Register(ParseOptions *po) {
       "Maximum number of sentences that we process at a time. "
       "This is to avoid OOM for very long input text. "
       "If you set it to -1, then we process all sentences in a single batch.");
+
+  po->Register("tts-silence-scale", &silence_scale,
+               "Duration of the pause is scaled by this number. So a smaller "
+               "value leads to a shorter pause.");
 }
 
 bool OfflineTtsConfig::Validate() const {
@@ -69,6 +140,11 @@ bool OfflineTtsConfig::Validate() const {
     }
   }
 
+  if (silence_scale < 0.001) {
+    SHERPA_ONNX_LOGE("--tts-silence-scale '%.3f' is too small", silence_scale);
+    return false;
+  }
+
   return model.Validate();
 }
 
@@ -79,7 +155,8 @@ std::string OfflineTtsConfig::ToString() const {
   os << "model=" << model.ToString() << ", ";
   os << "rule_fsts=\"" << rule_fsts << "\", ";
   os << "rule_fars=\"" << rule_fars << "\", ";
-  os << "max_num_sentences=" << max_num_sentences << ")";
+  os << "max_num_sentences=" << max_num_sentences << ", ";
+  os << "silence_scale=" << silence_scale << ")";
 
   return os.str();
 }

@@ -32,14 +32,20 @@ struct OfflineTtsConfig {
   // If you set it to -1, then we process all sentences in a single batch.
   int32_t max_num_sentences = 1;
 
+  // A silence interval containing audio samples with value close to 0.
+  //
+  // the duration of the new interval is old_duration * silence_scale.
+  float silence_scale = 0.2;
+
   OfflineTtsConfig() = default;
   OfflineTtsConfig(const OfflineTtsModelConfig &model,
                    const std::string &rule_fsts, const std::string &rule_fars,
-                   int32_t max_num_sentences)
+                   int32_t max_num_sentences, float silence_scale)
       : model(model),
         rule_fsts(rule_fsts),
         rule_fars(rule_fars),
-        max_num_sentences(max_num_sentences) {}
+        max_num_sentences(max_num_sentences),
+        silence_scale(silence_scale) {}
 
   void Register(ParseOptions *po);
   bool Validate() const;
@@ -50,6 +56,11 @@ struct OfflineTtsConfig {
 struct GeneratedAudio {
   std::vector<float> samples;
   int32_t sample_rate;
+
+  // Silence means pause here.
+  // If scale > 1, then it increases the duration of a pause
+  // If scale < 1, then it reduces the duration of a pause
+  GeneratedAudio ScaleSilence(float scale) const;
 };
 
 class OfflineTtsImpl;

@@ -7,12 +7,14 @@ public class OfflineTtsConfig {
     private final String ruleFsts;
     private final String ruleFars;
     private final int maxNumSentences;
+    private final float silenceScale;
 
     private OfflineTtsConfig(Builder builder) {
         this.model = builder.model;
         this.ruleFsts = builder.ruleFsts;
         this.ruleFars = builder.ruleFars;
         this.maxNumSentences = builder.maxNumSentences;
+        this.silenceScale = builder.silenceScale;
     }
 
     public static Builder builder() {
@@ -35,11 +37,16 @@ public class OfflineTtsConfig {
         return maxNumSentences;
     }
 
+    public float getSilenceScale() {
+        return silenceScale;
+    }
+
     public static class Builder {
         private OfflineTtsModelConfig model = OfflineTtsModelConfig.builder().build();
         private String ruleFsts = "";
         private String ruleFars = "";
         private int maxNumSentences = 1;
+        private float silenceScale = 0.2f;
 
         public OfflineTtsConfig build() {
             return new OfflineTtsConfig(this);
@@ -64,5 +71,10 @@ public class OfflineTtsConfig {
             this.maxNumSentences = maxNumSentences;
             return this;
         }
+
+        public Builder setSilenceScale(float silenceScale) {
+            this.silenceScale = silenceScale;
+            return this;
+        }
     }
 }

@@ -187,6 +187,9 @@ static OfflineTtsConfig GetOfflineTtsConfig(JNIEnv *env, jobject config) {
   fid = env->GetFieldID(cls, "maxNumSentences", "I");
   ans.max_num_sentences = env->GetIntField(config, fid);
 
+  fid = env->GetFieldID(cls, "silenceScale", "F");
+  ans.silence_scale = env->GetFloatField(config, fid);
+
   return ans;
 }
 

@@ -49,6 +49,7 @@ data class OfflineTtsConfig(
     var ruleFsts: String = "",
     var ruleFars: String = "",
     var maxNumSentences: Int = 1,
+    var silenceScale: Float = 0.2f,
 )
 
 class GeneratedAudio(

@@ -106,6 +106,7 @@ type
     RuleFsts: AnsiString;
     MaxNumSentences: Integer;
     RuleFars: AnsiString;
+    SilenceScale: Single;
 
     function ToString: AnsiString;
     class operator Initialize({$IFDEF FPC}var{$ELSE}out{$ENDIF} Dest: TSherpaOnnxOfflineTtsConfig);
@@ -777,6 +778,7 @@ type
     RuleFsts: PAnsiChar;
     MaxNumSentences: cint32;
     RuleFars: PAnsiChar;
+    SilenceScale: cfloat;
   end;
 
   PSherpaOnnxOfflineTtsConfig = ^SherpaOnnxOfflineTtsConfig;
@@ -1976,15 +1978,17 @@ begin
     'Model := %s, ' +
     'RuleFsts := %s, ' +
     'MaxNumSentences := %d, ' +
-    'RuleFars := %s' +
+    'RuleFars := %s, ' +
+    'SilenceScale := %f' +
     ')',
-    [Self.Model.ToString, Self.RuleFsts, Self.MaxNumSentences, Self.RuleFars
-    ]);
+    [Self.Model.ToString, Self.RuleFsts, Self.MaxNumSentences, Self.RuleFars,
+     Self.SilenceScale]);
 end;
 
 class operator TSherpaOnnxOfflineTtsConfig.Initialize({$IFDEF FPC}var{$ELSE}out{$ENDIF} Dest: TSherpaOnnxOfflineTtsConfig);
 begin
   Dest.MaxNumSentences := 1;
+  Dest.SilenceScale := 0.2;
 end;
 
 constructor TSherpaOnnxOfflineTts.Create(Config: TSherpaOnnxOfflineTtsConfig);
@@ -2027,6 +2031,7 @@ begin
   C.RuleFsts := PAnsiChar(Config.RuleFsts);
   C.MaxNumSentences := Config.MaxNumSentences;
   C.RuleFars := PAnsiChar(Config.RuleFars);
+  C.SilenceScale := Config.SilenceScale;
 
   Self.Handle := SherpaOnnxCreateOfflineTts(@C);
 

@@ -32,13 +32,15 @@ static void PybindOfflineTtsConfig(py::module *m) {
   py::class_<PyClass>(*m, "OfflineTtsConfig")
       .def(py::init<>())
       .def(py::init<const OfflineTtsModelConfig &, const std::string &,
-                    const std::string &, int32_t>(),
+                    const std::string &, int32_t, float>(),
            py::arg("model"), py::arg("rule_fsts") = "",
-           py::arg("rule_fars") = "", py::arg("max_num_sentences") = 2)
+           py::arg("rule_fars") = "", py::arg("max_num_sentences") = 2,
+           py::arg("silence_scale") = 0.2)
       .def_readwrite("model", &PyClass::model)
       .def_readwrite("rule_fsts", &PyClass::rule_fsts)
       .def_readwrite("rule_fars", &PyClass::rule_fars)
       .def_readwrite("max_num_sentences", &PyClass::max_num_sentences)
+      .def_readwrite("silence_scale", &PyClass::silence_scale)
       .def("validate", &PyClass::Validate)
       .def("__str__", &PyClass::ToString);
 }

@@ -804,13 +804,15 @@ func sherpaOnnxOfflineTtsConfig(
   model: SherpaOnnxOfflineTtsModelConfig,
   ruleFsts: String = "",
   ruleFars: String = "",
-  maxNumSentences: Int = 1
+  maxNumSentences: Int = 1,
+  silenceScale: Float = 0.2
 ) -> SherpaOnnxOfflineTtsConfig {
   return SherpaOnnxOfflineTtsConfig(
     model: model,
     rule_fsts: toCPointer(ruleFsts),
     max_num_sentences: Int32(maxNumSentences),
-    rule_fars: toCPointer(ruleFars)
+    rule_fars: toCPointer(ruleFars),
+    silence_scale: silenceScale
   )
 }
 

@@ -21,7 +21,7 @@ function freeConfig(config, Module) {
 
 // The user should free the returned pointers
 function initSherpaOnnxOfflineTtsVitsModelConfig(config, Module) {
-  const modelLen = Module.lengthBytesUTF8(config.model  || '')+ 1;
+  const modelLen = Module.lengthBytesUTF8(config.model || '') + 1;
   const lexiconLen = Module.lengthBytesUTF8(config.lexicon || '') + 1;
   const tokensLen = Module.lengthBytesUTF8(config.tokens || '') + 1;
   const dataDirLen = Module.lengthBytesUTF8(config.dataDir || '') + 1;
@@ -282,7 +282,7 @@ function initSherpaOnnxOfflineTtsModelConfig(config, Module) {
 function initSherpaOnnxOfflineTtsConfig(config, Module) {
   const modelConfig =
       initSherpaOnnxOfflineTtsModelConfig(config.offlineTtsModelConfig, Module);
-  const len = modelConfig.len + 3 * 4;
+  const len = modelConfig.len + 4 * 4;
   const ptr = Module._malloc(len);
 
   let offset = 0;
@@ -303,6 +303,10 @@ function initSherpaOnnxOfflineTtsConfig(config, Module) {
   offset += 4;
 
   Module.setValue(ptr + offset, buffer + ruleFstsLen, 'i8*');
+  offset += 4;
+
+  Module.setValue(ptr + offset, config.silenceScale || 0.2, 'float');
+  offset += 4;
 
   return {
     buffer: buffer, ptr: ptr, len: len, config: modelConfig,

@@ -22,7 +22,7 @@ static_assert(sizeof(SherpaOnnxOfflineTtsModelConfig) ==
                       sizeof(SherpaOnnxOfflineTtsKokoroModelConfig) + 3 * 4,
               "");
 static_assert(sizeof(SherpaOnnxOfflineTtsConfig) ==
-                  sizeof(SherpaOnnxOfflineTtsModelConfig) + 3 * 4,
+                  sizeof(SherpaOnnxOfflineTtsModelConfig) + 4 * 4,
               "");
 
 void MyPrint(SherpaOnnxOfflineTtsConfig *tts_config) {
@@ -68,6 +68,7 @@ void MyPrint(SherpaOnnxOfflineTtsConfig *tts_config) {
   fprintf(stdout, "rule_fsts: %s\n", tts_config->rule_fsts);
   fprintf(stdout, "rule_fars: %s\n", tts_config->rule_fars);
   fprintf(stdout, "max num sentences: %d\n", tts_config->max_num_sentences);
+  fprintf(stdout, "silence scale: %.3f\n", tts_config->silence_scale);
 }
 
 void CopyHeap(const char *src, int32_t num_bytes, char *dst) {