Text to speech API for Object Pascal. (#1273)

.github/workflows/pascal.yaml

@@ -119,13 +119,29 @@ jobs:
             cp -v install/lib/*.dll ../pascal-api-examples/vad
             cp -v install/lib/*.dll ../pascal-api-examples/vad-with-non-streaming-asr
 
-            cp -v ../sherpa-onnx/pascal-api/sherpa_onnx.pas ../pascal-api-examples/read-wav
-            cp -v ../sherpa-onnx/pascal-api/sherpa_onnx.pas ../pascal-api-examples/streaming-asr
-            cp -v ../sherpa-onnx/pascal-api/sherpa_onnx.pas ../pascal-api-examples/non-streaming-asr
-            cp -v ../sherpa-onnx/pascal-api/sherpa_onnx.pas ../pascal-api-examples/vad
-            cp -v ../sherpa-onnx/pascal-api/sherpa_onnx.pas ../pascal-api-examples/vad-with-non-streaming-asr
+            cp -v ../sherpa-onnx/pascal-api/*.pas ../pascal-api-examples/read-wav
+            cp -v ../sherpa-onnx/pascal-api/*.pas ../pascal-api-examples/streaming-asr
+            cp -v ../sherpa-onnx/pascal-api/*.pas ../pascal-api-examples/non-streaming-asr
+            cp -v ../sherpa-onnx/pascal-api/*.pas ../pascal-api-examples/vad
+            cp -v ../sherpa-onnx/pascal-api/*.pas ../pascal-api-examples/vad-with-non-streaming-asr
+            cp -v ../sherpa-onnx/pascal-api/*.pas ../pascal-api-examples/tts
           fi
 
+      - name:  Run Pascal test (TTS)
+        shell: bash
+        run: |
+          export PATH=/c/lazarus/fpc/3.2.2/bin/x86_64-win64:$PATH
+
+          cd ./pascal-api-examples
+          pushd tts
+
+          ./run-piper.sh
+          rm -rf vits-piper-*
+          ls -lh
+          echo "---"
+
+          popd
+
       - name:  Run Pascal test (VAD + non-streaming ASR)
         shell: bash
         run: |

pascal-api-examples/.gitignore

@@ -0,0 +1 @@
+link*.res

pascal-api-examples/README.md

@@ -13,3 +13,5 @@ https://k2-fsa.github.io/sherpa/onnx/pascal-api/index.html
 |[non-streaming-asr](./non-streaming-asr)| It shows how to use non-streaming models for speech recognition.|
 |[vad](./vad)| It shows how to use the voice activity detection API.|
 |[vad-with-non-streaming-asr](./vad-with-non-streaming-asr)| It shows how to use the voice activity detection API with non-streaming models for speech recognition.|
+|[portaudio-test](./portaudio-test)| It shows how to use PortAudio for recording and playing.|
+|[tts](./tts)| It shows how to use the text-to-speech API.|

pascal-api-examples/tts/.gitignore

@@ -0,0 +1,4 @@
+!run-*.sh
+piper
+piper-playback
+link*.res

pascal-api-examples/tts/README.md

@@ -0,0 +1,9 @@
+# Introduction
+
+This directory contains examples for how to use the TTS (text to speech) APIs.
+
+|Directory| Description|
+|---------|------------|
+|[run-piper.sh](./run-piper.sh)|It shows how to use models from [piper](https://github.com/rhasspy/piper) for text to speech.|
+|[run-piper-playback.sh](./run-piper-playback.sh)|It shows how to use models from [piper](https://github.com/rhasspy/piper) for text to speech. It plays the generated audio as it is still generating. |
+

pascal-api-examples/tts/piper-playback.pas

@@ -0,0 +1,238 @@
+{ Copyright (c)  2024  Xiaomi Corporation }
+program piper;
+{
+This file shows how to use the text to speech API of sherpa-onnx
+with Piper models.
+
+It generates speech from text and saves it to a wave file.
+
+Note that it plays the audio back as it is still generating.
+}
+
+{$mode objfpc}
+
+uses
+  {$ifdef unix}
+  cthreads,
+  {$endif}
+  SysUtils,
+  dos,
+  ctypes,
+  portaudio,
+  sherpa_onnx;
+
+var
+  CriticalSection: TRTLCriticalSection;
+
+  Tts: TSherpaOnnxOfflineTts;
+  Audio: TSherpaOnnxGeneratedAudio;
+  Resampler: TSherpaOnnxLinearResampler;
+
+  Text: AnsiString;
+  Speed: Single = 1.0;  {Use a larger value to speak faster}
+  SpeakerId: Integer = 0;
+  Buffer: TSherpaOnnxCircularBuffer;
+  FinishedGeneration: Boolean = False;
+  FinishedPlaying: Boolean = False;
+
+  Version: String;
+  EnvStr: String;
+  Status: Integer;
+  NumDevices: Integer;
+  DeviceIndex: Integer;
+  DeviceInfo: PPaDeviceInfo;
+
+  { If you get EDivByZero: Division by zero error, please change the sample rate
+    to the one supported by your microphone.
+  }
+  DeviceSampleRate: Integer = 48000;
+  I: Integer;
+  Param: TPaStreamParameters;
+  Stream: PPaStream;
+  Wave: TSherpaOnnxWave;
+
+function GenerateCallback(
+      Samples: pcfloat; N: cint32;
+      Arg: Pointer): cint; cdecl;
+begin
+  EnterCriticalSection(CriticalSection);
+  try
+    if Resampler <> nil then
+      Buffer.Push(Resampler.Resample(Samples, N, False))
+    else
+      Buffer.Push(Samples, N);
+  finally
+    LeaveCriticalSection(CriticalSection);
+  end;
+
+  { 1 means to continue generating; 0 means to stop generating. }
+  Result := 1;
+end;
+
+function PlayCallback(
+      input: Pointer; output: Pointer;
+      frameCount: culong;
+      timeInfo: PPaStreamCallbackTimeInfo;
+      statusFlags: TPaStreamCallbackFlags;
+      userData: Pointer ): cint; cdecl;
+var
+  Samples: TSherpaOnnxSamplesArray;
+  I: Integer;
+begin
+  EnterCriticalSection(CriticalSection);
+  try
+    if Buffer.Size >= frameCount then
+      begin
+        Samples := Buffer.Get(Buffer.Head, FrameCount);
+        Buffer.Pop(FrameCount);
+      end
+    else if Buffer.Size > 0 then
+      begin
+        Samples := Buffer.Get(Buffer.Head, Buffer.Size);
+        Buffer.Pop(Buffer.Size);
+        SetLength(Samples, frameCount);
+      end
+    else
+      SetLength(Samples, frameCount);
+
+    for I := 0 to frameCount - 1 do
+      pcfloat(output)[I] := Samples[I];
+
+    if (Buffer.Size > 0) or (not FinishedGeneration) then
+      Result := paContinue
+    else
+      begin
+        Result := paComplete;
+        FinishedPlaying := True;
+      end;
+  finally
+    LeaveCriticalSection(CriticalSection);
+  end;
+end;
+
+function GetOfflineTts: TSherpaOnnxOfflineTts;
+var
+  Config: TSherpaOnnxOfflineTtsConfig;
+begin
+  Config.Model.Vits.Model := './vits-piper-en_US-libritts_r-medium/en_US-libritts_r-medium.onnx';
+  Config.Model.Vits.Tokens := './vits-piper-en_US-libritts_r-medium/tokens.txt';
+  Config.Model.Vits.DataDir := './vits-piper-en_US-libritts_r-medium/espeak-ng-data';
+  Config.Model.NumThreads := 1;
+  Config.Model.Debug := False;
+  Config.MaxNumSentences := 1;
+
+  Result := TSherpaOnnxOfflineTts.Create(Config);
+end;
+
+begin
+  Tts := GetOfflineTts;
+  if Tts.GetSampleRate <> DeviceSampleRate then
+    Resampler := TSherpaOnnxLinearResampler.Create(Tts.GetSampleRate, DeviceSampleRate);
+
+  Version := String(Pa_GetVersionText);
+  WriteLn('Version is ', Version);
+  Status := Pa_Initialize;
+  if Status <> paNoError then
+    begin
+      WriteLn('Failed to initialize portaudio, ', Pa_GetErrorText(Status));
+      Exit;
+    end;
+
+  NumDevices := Pa_GetDeviceCount;
+  WriteLn('Num devices: ', NumDevices);
+
+  DeviceIndex := Pa_GetDefaultOutputDevice;
+
+  if DeviceIndex = paNoDevice then
+    begin
+      WriteLn('No default output device found');
+      Pa_Terminate;
+      Exit;
+    end;
+
+  EnvStr := GetEnv('SHERPA_ONNX_MIC_DEVICE');
+  if EnvStr <> '' then
+    begin
+      DeviceIndex := StrToIntDef(EnvStr, DeviceIndex);
+      WriteLn('Use device index from environment variable SHERPA_ONNX_MIC_DEVICE: ', EnvStr);
+    end;
+
+  for I := 0 to (NumDevices - 1) do
+    begin
+      DeviceInfo := Pa_GetDeviceInfo(I);
+      if I = DeviceIndex then
+        { WriteLn(Format(' * %d %s', [I, DeviceInfo^.Name])) }
+        WriteLn(Format(' * %d %s', [I, AnsiString(DeviceInfo^.Name)]))
+      else
+        WriteLn(Format('   %d %s', [I, AnsiString(DeviceInfo^.Name)]));
+    end;
+
+  WriteLn('Use device ', DeviceIndex);
+  WriteLn(' Name ', Pa_GetDeviceInfo(DeviceIndex)^.Name);
+  WriteLn(' Max output channels ', Pa_GetDeviceInfo(DeviceIndex)^.MaxOutputChannels);
+
+  Initialize(Param);
+  Param.Device := DeviceIndex;
+  Param.ChannelCount := 1;
+  Param.SampleFormat := paFloat32;
+  param.SuggestedLatency := Pa_GetDeviceInfo(DeviceIndex)^.DefaultHighOutputLatency;
+  param.HostApiSpecificStreamInfo := nil;
+
+  Buffer := TSherpaOnnxCircularBuffer.Create(30 * DeviceSampleRate);
+
+
+  { Note(fangjun): PortAudio invokes PlayCallback in a separate thread. }
+  Status := Pa_OpenStream(stream, nil, @Param, DeviceSampleRate, paFramesPerBufferUnspecified, paNoFlag,
+    PPaStreamCallback(@PlayCallback), nil);
+
+  if Status <> paNoError then
+    begin
+      WriteLn('Failed to open stream, ', Pa_GetErrorText(Status));
+      Pa_Terminate;
+      Exit;
+    end;
+
+  InitCriticalSection(CriticalSection);
+
+  Status := Pa_StartStream(stream);
+  if Status <> paNoError then
+    begin
+      WriteLn('Failed to start stream, ', Pa_GetErrorText(Status));
+      Pa_Terminate;
+      Exit;
+    end;
+
+  WriteLn('There are ', Tts.GetNumSpeakers, ' speakers');
+
+  Text := 'Today as always, men fall into two groups: slaves and free men. Whoever does not have two-thirds of his day for himself, is a slave, whatever he may be: a statesman, a businessman, an official, or a scholar.';
+
+  Audio :=  Tts.Generate(Text, SpeakerId, Speed,
+    PSherpaOnnxGeneratedAudioCallbackWithArg(@GenerateCallback), nil);
+  FinishedGeneration := True;
+  SherpaOnnxWriteWave('./libritts_r-generated.wav', Audio.Samples, Audio.SampleRate);
+  WriteLn('Saved to ./libritts_r-generated.wav');
+
+  while not FinishedPlaying do
+    Pa_Sleep(100);  {sleep for 0.1 second }
+    {TODO(fangjun): Use an event to indicate the play is finished}
+
+  DoneCriticalSection(CriticalSection);
+
+  FreeAndNil(Tts);
+  FreeAndNil(Resampler);
+
+  Status := Pa_CloseStream(stream);
+  if Status <> paNoError then
+    begin
+      WriteLn('Failed to close stream, ', Pa_GetErrorText(Status));
+      Exit;
+    end;
+
+  Status := Pa_Terminate;
+  if Status <> paNoError then
+    begin
+      WriteLn('Failed to deinitialize portaudio, ', Pa_GetErrorText(Status));
+      Exit;
+    end;
+end.
+

pascal-api-examples/tts/piper.pas

@@ -0,0 +1,54 @@
+{ Copyright (c)  2024  Xiaomi Corporation }
+program piper;
+{
+This file shows how to use the text to speech API of sherpa-onnx
+with Piper models.
+
+It generates speech from text and saves it to a wave file.
+
+If you want to play it while it is generating, please see
+./piper-playback.pas
+}
+
+{$mode objfpc}
+
+uses
+  SysUtils,
+  sherpa_onnx;
+
+function GetOfflineTts: TSherpaOnnxOfflineTts;
+var
+  Config: TSherpaOnnxOfflineTtsConfig;
+begin
+  Config.Model.Vits.Model := './vits-piper-en_US-libritts_r-medium/en_US-libritts_r-medium.onnx';
+  Config.Model.Vits.Tokens := './vits-piper-en_US-libritts_r-medium/tokens.txt';
+  Config.Model.Vits.DataDir := './vits-piper-en_US-libritts_r-medium/espeak-ng-data';
+  Config.Model.NumThreads := 1;
+  Config.Model.Debug := False;
+  Config.MaxNumSentences := 1;
+
+  Result := TSherpaOnnxOfflineTts.Create(Config);
+end;
+
+var
+  Tts: TSherpaOnnxOfflineTts;
+  Audio: TSherpaOnnxGeneratedAudio;
+
+  Text: AnsiString;
+  Speed: Single = 1.0;  {Use a larger value to speak faster}
+  SpeakerId: Integer = 0;
+
+begin
+  Tts := GetOfflineTts;
+
+  WriteLn('There are ', Tts.GetNumSpeakers, ' speakers');
+
+  Text := 'Today as always, men fall into two groups: slaves and free men. Whoever does not have two-thirds of his day for himself, is a slave, whatever he may be: a statesman, a businessman, an official, or a scholar.';
+
+  Audio :=  Tts.Generate(Text, SpeakerId, Speed);
+  SherpaOnnxWriteWave('./libritts_r-generated.wav', Audio.Samples, Audio.SampleRate);
+  WriteLn('Saved to ./libritts_r-generated.wav');
+
+  FreeAndNil(Tts);
+end.
+

pascal-api-examples/tts/run-piper-playback.sh

@@ -0,0 +1,45 @@
+#!/usr/bin/env bash
+
+set -ex
+
+SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
+SHERPA_ONNX_DIR=$(cd $SCRIPT_DIR/../.. && pwd)
+
+echo "SHERPA_ONNX_DIR: $SHERPA_ONNX_DIR"
+
+if [[ ! -f ../../build/install/lib/libsherpa-onnx-c-api.dylib  && ! -f ../../build/install/lib/libsherpa-onnx-c-api.so && ! -f ../../build/install/lib/sherpa-onnx-c-api.dll ]]; then
+  mkdir -p ../../build
+  pushd ../../build
+  cmake \
+    -DCMAKE_INSTALL_PREFIX=./install \
+    -DSHERPA_ONNX_ENABLE_PYTHON=OFF \
+    -DSHERPA_ONNX_ENABLE_TESTS=OFF \
+    -DSHERPA_ONNX_ENABLE_CHECK=OFF \
+    -DBUILD_SHARED_LIBS=ON \
+    -DSHERPA_ONNX_ENABLE_PORTAUDIO=OFF \
+    ..
+
+  cmake --build . --target install --config Release
+  popd
+fi
+
+if [[ ! -f ./vits-piper-en_US-libritts_r-medium/tokens.txt ]]; then
+  curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/tts-models/vits-piper-en_US-libritts_r-medium.tar.bz2
+  tar xf vits-piper-en_US-libritts_r-medium.tar.bz2
+  rm vits-piper-en_US-libritts_r-medium.tar.bz2
+fi
+
+fpc \
+  -dSHERPA_ONNX_USE_SHARED_LIBS \
+  -Fu$SHERPA_ONNX_DIR/sherpa-onnx/pascal-api \
+  -Fl$SHERPA_ONNX_DIR/build/install/lib \
+  -Fl/usr/local/Cellar/portaudio/19.7.0/lib \
+  ./piper-playback.pas
+
+# Please see ../portaudio-test/README.md
+# for how to install portaudio on macOS
+
+export LD_LIBRARY_PATH=$SHERPA_ONNX_DIR/build/install/lib:$LD_LIBRARY_PATH
+export DYLD_LIBRARY_PATH=$SHERPA_ONNX_DIR/build/install/lib:$DYLD_LIBRARY_PATH
+
+./piper-playback