Using patched espeak-ng

.gitignore

@@ -5,6 +5,7 @@ tmp/
 
 *.py[cod]
 *.egg
+*.egg-info/
 build
 htmlcov
 

Dockerfile

@@ -14,8 +14,8 @@ ENV DEBIAN_FRONTEND=noninteractive
 WORKDIR /build
 
 # Build minimal version of espeak-ng
-ADD lib/espeak-ng-1.51.tar.gz ./
-RUN cd espeak-ng-1.51 && \
+ADD lib/espeak-ng-1.52-patched.tar.gz ./
+RUN cd espeak-ng && \
     ./autogen.sh && \
     ./configure \
         --without-pcaudiolib \
@@ -30,9 +30,10 @@ RUN cd espeak-ng-1.51 && \
     make install
 
 # Copy onnxruntime library
-COPY lib/ ./lib/
-RUN mkdir -p /usr/local/include/onnxruntime && \
-    tar -C /usr/local/include/onnxruntime \
+COPY lib/onnxruntime-linux-*.tgz ./lib/
+RUN export ONNX_DIR="./lib/Linux-$(uname -m)" && \
+    mkdir -p "${ONNX_DIR}" && \
+    tar -C "${ONNX_DIR}" \
         --strip-components 1 \
         -xvf "lib/onnxruntime-linux-${TARGETARCH}${TARGETVARIANT}.tgz"
 
@@ -49,7 +50,7 @@ WORKDIR /dist
 RUN mkdir -p piper && \
     cp -d /usr/lib64/libespeak-ng.so* ./piper/ && \
     cp -dR /usr/share/espeak-ng-data ./piper/ && \
-    cp -d /usr/local/include/onnxruntime/lib/libonnxruntime.so.* ./piper/ && \
+    find /build/lib/ -name 'libonnxruntime.so.*' -exec cp -d {} ./piper/ \; && \
     cp /build/build/piper ./piper/ && \
     tar -czf "piper_${TARGETARCH}${TARGETVARIANT}.tar.gz" piper/
 

Dockerfile.dockerignore

@@ -1,4 +1,5 @@
 *
 !Makefile
 !src/cpp/
-!lib/
+!lib/onnxruntime*.tgz
+!lib/espeak-ng*.tar.gz

README.md

@@ -39,7 +39,9 @@ Download a release:
 * [amd64](https://github.com/rhasspy/piper/releases/download/v0.0.2/piper_amd64.tar.gz) (desktop Linux)
 * [arm64](https://github.com/rhasspy/piper/releases/download/v0.0.2/piper_arm64.tar.gz) (Raspberry Pi 4)
 
-If you want to build from source, see the [Makefile](Makefile) and [C++ source](src/cpp). Last tested with [onnxruntime](https://github.com/microsoft/onnxruntime) 1.13.1.
+If you want to build from source, see the [Makefile](Makefile) and [C++ source](src/cpp). Piper depends on a patched `espeak-ng` in [lib](lib).
+
+Last tested with [onnxruntime](https://github.com/microsoft/onnxruntime) 1.14.1.
 
 
 ## Usage

lib/.gitignore

@@ -0,0 +1,2 @@
+espeak-ng/
+Linux-*/

src/cpp/config.hpp

@@ -32,8 +32,10 @@ struct eSpeakConfig {
   // Characters that eSpeak uses to break apart paragraphs/sentences
   set<Phoneme> clauseBreakers{U'.', U'?', U'!', U',', U';', U':'};
 
-  // Characters that piper will use to split utterances
-  set<Phoneme> sentenceBreakers{U'.', U'?', U'!'};
+  Phoneme fullStop = U'.';
+  Phoneme comma = U',';
+  Phoneme question = U'?';
+  Phoneme exclamation = U'!';
 };
 
 struct PhonemizeConfig {

src/cpp/main.cpp

@@ -53,7 +53,8 @@ int main(int argc, char *argv[]) {
   RunConfig runConfig;
   parseArgs(argc, argv, runConfig);
 
-  auto exePath = filesystem::path(argv[0]);
+  // NOTE: This won't work for Windows (need GetModuleFileName)
+  auto exePath = filesystem::canonical("/proc/self/exe");
   piper::initialize(exePath.parent_path());
 
   piper::Voice voice;

src/cpp/phonemize.hpp

@@ -15,6 +15,13 @@
 #include "config.hpp"
 #include "utf8.h"
 
+#define CLAUSE_INTONATION_FULL_STOP   0x00000000
+#define CLAUSE_INTONATION_COMMA       0x00001000
+#define CLAUSE_INTONATION_QUESTION    0x00002000
+#define CLAUSE_INTONATION_EXCLAMATION 0x00003000
+
+#define CLAUSE_TYPE_SENTENCE          0x00080000
+
 using namespace std;
 
 namespace piper {
@@ -54,13 +61,15 @@ void phonemize(string text, PhonemizeConfig &phonemizeConfig,
 
   vector<Phoneme> *sentencePhonemes = nullptr;
   const char *inputTextPointer = textCopy.c_str();
-  size_t clauseBreakerIndex = 0;
+  int terminator = 0;
 
   while (inputTextPointer != NULL) {
+    // Modified espeak-ng API to get access to clause terminator
     string clausePhonemes(
-        espeak_TextToPhonemes((const void **)&inputTextPointer,
+        espeak_TextToPhonemes2((const void **)&inputTextPointer,
                               /*textmode*/ espeakCHARS_AUTO,
-                              /*phonememode = IPA*/ 0x02));
+                              /*phonememode = IPA*/ 0x02,
+                               &terminator));
 
     utf8::iterator phonemeIter(clausePhonemes.begin(), clausePhonemes.begin(),
                                clausePhonemes.end());
@@ -74,17 +83,25 @@ void phonemize(string text, PhonemizeConfig &phonemizeConfig,
     }
 
     sentencePhonemes->insert(sentencePhonemes->end(), phonemeIter, phonemeEnd);
-    if (clauseBreakerIndex < textClauseBreakers.size()) {
-      auto clauseBreaker = textClauseBreakers[clauseBreakerIndex];
-      sentencePhonemes->push_back(clauseBreaker);
-      if (phonemizeConfig.eSpeak->sentenceBreakers.contains(clauseBreaker)) {
+
+    // Add appropriate puntuation depending on terminator type
+    int intonation = terminator & 0x0000F000;
+    if (intonation == CLAUSE_INTONATION_FULL_STOP) {
+      sentencePhonemes->push_back(phonemizeConfig.eSpeak->fullStop);
+    } else if (intonation == CLAUSE_INTONATION_COMMA) {
+      sentencePhonemes->push_back(phonemizeConfig.eSpeak->comma);
+    } else if (intonation == CLAUSE_INTONATION_QUESTION) {
+      sentencePhonemes->push_back(phonemizeConfig.eSpeak->question);
+    } else if (intonation == CLAUSE_INTONATION_EXCLAMATION) {
+      sentencePhonemes->push_back(phonemizeConfig.eSpeak->exclamation);
+    }
+
+    if ((terminator & CLAUSE_TYPE_SENTENCE) == CLAUSE_TYPE_SENTENCE) {
         // End of sentence
         sentencePhonemes = nullptr;
-      }
-
-      clauseBreakerIndex++;
     }
-  }
+
+  }  // while inputTextPointer != NULL
 
 } /* phonemize */
 

src/cpp/piper.hpp

@@ -30,7 +30,7 @@ struct Voice {
 void initialize(std::filesystem::path cwd) {
   const char *dataPath = NULL;
 
-  auto cwdDataPath = cwd.append("espeak-ng-data");
+  auto cwdDataPath = std::filesystem::absolute(cwd.append("espeak-ng-data"));
   if (std::filesystem::is_directory(cwdDataPath)) {
     dataPath = cwdDataPath.c_str();
   }