Add Pascal API for Dolphin CTC models (#2096)

.github/workflows/pascal.yaml

@@ -149,6 +149,11 @@ jobs:
           cd ./pascal-api-examples
 
           pushd non-streaming-asr
+
+          ./run-dolphin-ctc.sh
+          rm -rf sherpa-onnx-*
+          echo "---"
+
           ./run-zipformer-transducer.sh
           rm -rf sherpa-onnx-*
           echo "---"
@@ -253,7 +258,13 @@ jobs:
 
           cd ./pascal-api-examples
 
+
           pushd vad-with-non-streaming-asr
+
+          time ./run-vad-with-dolphin-ctc.sh
+          rm -rf sherpa-onnx-*
+          echo "---"
+
           time ./run-vad-with-moonshine.sh
           rm -rf sherpa-onnx-*
           echo "---"

README.md

@@ -60,7 +60,7 @@ This repository supports running the following functions **locally**
 
 on the following platforms and operating systems:
 
-  - x86, ``x86_64``, 32-bit ARM, 64-bit ARM (arm64, aarch64), RISC-V (riscv64)
+  - x86, ``x86_64``, 32-bit ARM, 64-bit ARM (arm64, aarch64), RISC-V (riscv64), **RK NPU**
   - Linux, macOS, Windows, openKylin
   - Android, WearOS
   - iOS

pascal-api-examples/non-streaming-asr/README.md

@@ -5,6 +5,7 @@ APIs with non-streaming models for speech recognition.
 
 |File|Description|
 |----|-----------|
+|[run-dolphin-ctc.sh](./run-dolphin-ctc.sh)|Use a non-streaming [Dolphin](https://github.com/DataoceanAI/Dolphin) CTC model for speech recognition|
 |[run-nemo-ctc.sh](./run-nemo-ctc.sh)|Use a non-streaming NeMo CTC model for speech recognition|
 |[run-nemo-transducer.sh](./run-nemo-transducer.sh)|Use a non-streaming NeMo transducer model for speech recognition|
 |[run-paraformer-itn.sh](./run-paraformer-itn.sh)|Use a non-streaming Paraformer model for speech recognition with inverse text normalization for numbers|

pascal-api-examples/non-streaming-asr/dolphin_ctc.pas

@@ -0,0 +1,76 @@
+{ Copyright (c)  2025  Xiaomi Corporation }
+
+{
+This file shows how to use a non-streaming Dolphin CTC model
+to decode files.
+
+You can download the model files from
+https://github.com/k2-fsa/sherpa-onnx/releases/tag/asr-models
+}
+
+program dolphin_ctc;
+
+{$mode objfpc}
+
+uses
+  sherpa_onnx,
+  DateUtils,
+  SysUtils;
+
+var
+  Wave: TSherpaOnnxWave;
+  WaveFilename: AnsiString;
+
+  Config: TSherpaOnnxOfflineRecognizerConfig;
+  Recognizer: TSherpaOnnxOfflineRecognizer;
+  Stream: TSherpaOnnxOfflineStream;
+  RecognitionResult: TSherpaOnnxOfflineRecognizerResult;
+
+  Start: TDateTime;
+  Stop: TDateTime;
+
+  Elapsed: Single;
+  Duration: Single;
+  RealTimeFactor: Single;
+begin
+  Initialize(Config);
+
+  Config.ModelConfig.Dolphin.Model := './sherpa-onnx-dolphin-base-ctc-multi-lang-int8-2025-04-02/model.int8.onnx';
+  Config.ModelConfig.Tokens := './sherpa-onnx-dolphin-base-ctc-multi-lang-int8-2025-04-02/tokens.txt';
+  Config.ModelConfig.Provider := 'cpu';
+  Config.ModelConfig.NumThreads := 1;
+  Config.ModelConfig.Debug := False;
+
+  WaveFilename := './sherpa-onnx-dolphin-base-ctc-multi-lang-int8-2025-04-02/test_wavs/0.wav';
+
+  Wave := SherpaOnnxReadWave(WaveFilename);
+
+  Recognizer := TSherpaOnnxOfflineRecognizer.Create(Config);
+  Stream := Recognizer.CreateStream();
+  Start := Now;
+
+  Stream.AcceptWaveform(Wave.Samples, Wave.SampleRate);
+  Recognizer.Decode(Stream);
+
+  RecognitionResult := Recognizer.GetResult(Stream);
+
+  Stop := Now;
+
+  Elapsed := MilliSecondsBetween(Stop, Start) / 1000;
+  Duration := Length(Wave.Samples) / Wave.SampleRate;
+  RealTimeFactor := Elapsed / Duration;
+
+  WriteLn(RecognitionResult.ToString);
+  WriteLn(Format('NumThreads %d', [Config.ModelConfig.NumThreads]));
+  WriteLn(Format('Elapsed %.3f s', [Elapsed]));
+  WriteLn(Format('Wave duration %.3f s', [Duration]));
+  WriteLn(Format('RTF = %.3f/%.3f = %.3f', [Elapsed, Duration, RealTimeFactor]));
+
+  {Free resources to avoid memory leak.
+
+  Note: You don't need to invoke them for this simple script.
+  However, you have to invoke them in your own large/complex project.
+  }
+  FreeAndNil(Stream);
+  FreeAndNil(Recognizer);
+end.

pascal-api-examples/non-streaming-asr/run-dolphin-ctc.sh

@@ -0,0 +1,42 @@
+#!/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: $SHERPA_ONNX_DIR"
+
+if [[ ! -f ../../build/install/lib/libsherpa-onnx-c-api.dylib  && ! -f ../../build/install/lib/libsherpa-onnx-c-api.so && ! -f ../../build/install/lib/sherpa-onnx-c-api.dll ]]; then
+  mkdir -p ../../build
+  pushd ../../build
+  cmake \
+    -DCMAKE_INSTALL_PREFIX=./install \
+    -DSHERPA_ONNX_ENABLE_PYTHON=OFF \
+    -DSHERPA_ONNX_ENABLE_TESTS=OFF \
+    -DSHERPA_ONNX_ENABLE_CHECK=OFF \
+    -DBUILD_SHARED_LIBS=ON \
+    -DSHERPA_ONNX_ENABLE_PORTAUDIO=OFF \
+    ..
+
+  cmake --build . --target install --config Release
+  ls -lh lib
+  popd
+fi
+
+if [ ! -f ./sherpa-onnx-dolphin-base-ctc-multi-lang-int8-2025-04-02/model.int8.onnx ]; then
+  curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-dolphin-base-ctc-multi-lang-int8-2025-04-02.tar.bz2
+  tar xvf sherpa-onnx-dolphin-base-ctc-multi-lang-int8-2025-04-02.tar.bz2
+  rm sherpa-onnx-dolphin-base-ctc-multi-lang-int8-2025-04-02.tar.bz2
+fi
+
+fpc \
+  -dSHERPA_ONNX_USE_SHARED_LIBS \
+  -Fu$SHERPA_ONNX_DIR/sherpa-onnx/pascal-api \
+  -Fl$SHERPA_ONNX_DIR/build/install/lib \
+  ./dolphin_ctc.pas
+
+export LD_LIBRARY_PATH=$SHERPA_ONNX_DIR/build/install/lib:$LD_LIBRARY_PATH
+export DYLD_LIBRARY_PATH=$SHERPA_ONNX_DIR/build/install/lib:$DYLD_LIBRARY_PATH
+
+./dolphin_ctc

pascal-api-examples/vad-with-non-streaming-asr/README.md

@@ -6,7 +6,10 @@ with non-streaming speech recognition models.
 
 |Directory| Description|
 |---------|------------|
-|[run-vad-with-whisper.sh](./run-vad-with-whisper.sh)|It shows how to use the VAD + Whisper for speech recognition.|
+|[run-vad-with-dolphin-ctc.sh](./run-vad-with-dolphin-ctc.sh)|It shows how to use the VAD + [Dolphin](https://github.com/DataoceanAI/Dolphin) for speech recognition.|
-|[run-vad-with-sense-voice.sh](./run-vad-with-sense-voice.sh)|It shows how to use the VAD + SenseVoice for speech recognition.|
+|[run-vad-with-whisper.sh](./run-vad-with-whisper.sh)|It shows how to use the VAD + [Whisper](https://github.com/openai/whisper) for speech recognition.|
+|[run-vad-with-sense-voice.sh](./run-vad-with-sense-voice.sh)|It shows how to use the VAD + [SenseVoice](https://github.com/FunAudioLLM/SenseVoice) for speech recognition.|
+|[run-vad-with-moonshine.sh](./run-vad-with-moonshine.sh)|It shows how to use the VAD + [Moonshine](https://github.com/usefulsensors/moonshine) for speech recognition.|
+
 
 Please refer to [non-streaming-asr](../non-streaming-asr) for more kinds of non-streaming models.

pascal-api-examples/vad-with-non-streaming-asr/run-vad-with-dolphin-ctc.sh

@@ -0,0 +1,49 @@
+#!/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: $SHERPA_ONNX_DIR"
+
+if [[ ! -f ../../build/install/lib/libsherpa-onnx-c-api.dylib  && ! -f ../../build/install/lib/libsherpa-onnx-c-api.so && ! -f ../../build/install/lib/sherpa-onnx-c-api.dll ]]; then
+  mkdir -p ../../build
+  pushd ../../build
+  cmake \
+    -DCMAKE_INSTALL_PREFIX=./install \
+    -DSHERPA_ONNX_ENABLE_PYTHON=OFF \
+    -DSHERPA_ONNX_ENABLE_TESTS=OFF \
+    -DSHERPA_ONNX_ENABLE_CHECK=OFF \
+    -DBUILD_SHARED_LIBS=ON \
+    -DSHERPA_ONNX_ENABLE_PORTAUDIO=OFF \
+    ..
+
+  cmake --build . --target install --config Release
+  popd
+fi
+
+if [[ ! -f ./silero_vad.onnx ]]; then
+  curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/silero_vad.onnx
+fi
+
+if [ ! -f ./lei-jun-test.wav ]; then
+  curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/lei-jun-test.wav
+fi
+
+if [ ! -f ./sherpa-onnx-dolphin-base-ctc-multi-lang-int8-2025-04-02/model.int8.onnx ]; then
+  curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-dolphin-base-ctc-multi-lang-int8-2025-04-02.tar.bz2
+  tar xvf sherpa-onnx-dolphin-base-ctc-multi-lang-int8-2025-04-02.tar.bz2
+  rm sherpa-onnx-dolphin-base-ctc-multi-lang-int8-2025-04-02.tar.bz2
+fi
+
+fpc \
+  -dSHERPA_ONNX_USE_SHARED_LIBS \
+  -Fu$SHERPA_ONNX_DIR/sherpa-onnx/pascal-api \
+  -Fl$SHERPA_ONNX_DIR/build/install/lib \
+  ./vad_with_dolphin.pas
+
+export LD_LIBRARY_PATH=$SHERPA_ONNX_DIR/build/install/lib:$LD_LIBRARY_PATH
+export DYLD_LIBRARY_PATH=$SHERPA_ONNX_DIR/build/install/lib:$DYLD_LIBRARY_PATH
+
+./vad_with_dolphin

pascal-api-examples/vad-with-non-streaming-asr/vad_with_dolphin.pas

@@ -0,0 +1,135 @@
+{ Copyright (c)  2025  Xiaomi Corporation }
+
+{
+This file shows how to use a non-streaming Dolphin model
+with silero VAD to decode files.
+
+You can download the model files from
+https://github.com/k2-fsa/sherpa-onnx/releases/tag/asr-models
+}
+
+program vad_with_dolphin;
+
+{$mode objfpc}
+
+uses
+  sherpa_onnx,
+  SysUtils;
+
+function CreateVad(): TSherpaOnnxVoiceActivityDetector;
+var
+  Config: TSherpaOnnxVadModelConfig;
+
+  SampleRate: Integer;
+  WindowSize: Integer;
+begin
+  Initialize(Config);
+
+  SampleRate := 16000; {Please don't change it unless you know the details}
+  WindowSize := 512; {Please don't change it unless you know the details}
+
+  Config.SileroVad.Model := './silero_vad.onnx';
+  Config.SileroVad.MinSpeechDuration := 0.5;
+  Config.SileroVad.MinSilenceDuration := 0.5;
+  Config.SileroVad.Threshold := 0.5;
+  Config.SileroVad.WindowSize := WindowSize;
+  Config.NumThreads:= 1;
+  Config.Debug:= True;
+  Config.Provider:= 'cpu';
+  Config.SampleRate := SampleRate;
+
+  Result := TSherpaOnnxVoiceActivityDetector.Create(Config, 30);
+end;
+
+function CreateOfflineRecognizer(): TSherpaOnnxOfflineRecognizer;
+var
+  Config: TSherpaOnnxOfflineRecognizerConfig;
+begin
+  Initialize(Config);
+
+  Config.ModelConfig.Dolphin.Model := './sherpa-onnx-dolphin-base-ctc-multi-lang-int8-2025-04-02/model.int8.onnx';
+  Config.ModelConfig.Tokens := './sherpa-onnx-dolphin-base-ctc-multi-lang-int8-2025-04-02/tokens.txt';
+  Config.ModelConfig.Provider := 'cpu';
+  Config.ModelConfig.NumThreads := 1;
+  Config.ModelConfig.Debug := False;
+
+  Result := TSherpaOnnxOfflineRecognizer.Create(Config);
+end;
+
+var
+  Wave: TSherpaOnnxWave;
+
+  Recognizer: TSherpaOnnxOfflineRecognizer;
+  Vad: TSherpaOnnxVoiceActivityDetector;
+
+  Offset: Integer;
+  WindowSize: Integer;
+  SpeechSegment: TSherpaOnnxSpeechSegment;
+
+  Start: Single;
+  Duration: Single;
+
+  Stream: TSherpaOnnxOfflineStream;
+  RecognitionResult: TSherpaOnnxOfflineRecognizerResult;
+begin
+  Vad := CreateVad();
+  Recognizer := CreateOfflineRecognizer();
+
+  Wave := SherpaOnnxReadWave('./lei-jun-test.wav');
+  if Wave.SampleRate <> Vad.Config.SampleRate then
+    begin
+      WriteLn(Format('Expected sample rate: %d. Given: %d',
+        [Vad.Config.SampleRate, Wave.SampleRate]));
+
+      Exit;
+    end;
+
+  WindowSize := Vad.Config.SileroVad.WindowSize;
+  Offset := 0;
+  while Offset + WindowSize <= Length(Wave.Samples) do
+    begin
+      Vad.AcceptWaveform(Wave.Samples, Offset, WindowSize);
+      Offset += WindowSize;
+
+      while not Vad.IsEmpty do
+        begin
+          SpeechSegment := Vad.Front();
+          Vad.Pop();
+          Stream := Recognizer.CreateStream();
+
+          Stream.AcceptWaveform(SpeechSegment.Samples, Wave.SampleRate);
+          Recognizer.Decode(Stream);
+          RecognitionResult := Recognizer.GetResult(Stream);
+
+          Start := SpeechSegment.Start / Wave.SampleRate;
+          Duration := Length(SpeechSegment.Samples) / Wave.SampleRate;
+          WriteLn(Format('%.3f -- %.3f %s',
+            [Start, Start + Duration, RecognitionResult.Text]));
+
+          FreeAndNil(Stream);
+        end;
+    end;
+
+  Vad.Flush;
+
+  while not Vad.IsEmpty do
+    begin
+      SpeechSegment := Vad.Front();
+      Vad.Pop();
+      Stream := Recognizer.CreateStream();
+
+      Stream.AcceptWaveform(SpeechSegment.Samples, Wave.SampleRate);
+      Recognizer.Decode(Stream);
+      RecognitionResult := Recognizer.GetResult(Stream);
+
+      Start := SpeechSegment.Start / Wave.SampleRate;
+      Duration := Length(SpeechSegment.Samples) / Wave.SampleRate;
+      WriteLn(Format('%.3f -- %.3f %s',
+        [Start, Start + Duration, RecognitionResult.Text]));
+
+      FreeAndNil(Stream);
+    end;
+
+  FreeAndNil(Recognizer);
+  FreeAndNil(Vad);
+end.

pascal-api-examples/vad-with-non-streaming-asr/vad_with_sense_voice.pas

@@ -8,7 +8,7 @@ You can download the model files from
 https://github.com/k2-fsa/sherpa-onnx/releases/tag/asr-models
 }
 
-program vad_with_whisper;
+program vad_with_sense_voice;
 
 {$mode objfpc}
 

sherpa-onnx/c-api/c-api.cc

@@ -1969,7 +1969,7 @@ int32_t SherpaOnnxLinearResamplerResampleGetOutputSampleRate(
   return p->impl->GetOutputSamplingRate();
 }
 
-void SherpaOnnxLinearResamplerReset(SherpaOnnxLinearResampler *p) {
+void SherpaOnnxLinearResamplerReset(const SherpaOnnxLinearResampler *p) {
   p->impl->Reset();
 }
 

sherpa-onnx/pascal-api/sherpa_onnx.pas

@@ -270,6 +270,11 @@ type
     function ToString: AnsiString;
   end;
 
+  TSherpaOnnxOfflineDolphinModelConfig = record
+    Model: AnsiString;
+    function ToString: AnsiString;
+  end;
+
   TSherpaOnnxOfflineWhisperModelConfig = record
     Encoder: AnsiString;
     Decoder: AnsiString;
@@ -331,6 +336,7 @@ type
     SenseVoice: TSherpaOnnxOfflineSenseVoiceModelConfig;
     Moonshine: TSherpaOnnxOfflineMoonshineModelConfig;
     FireRedAsr: TSherpaOnnxOfflineFireRedAsrModelConfig;
+    Dolphin: TSherpaOnnxOfflineDolphinModelConfig;
     class operator Initialize({$IFDEF FPC}var{$ELSE}out{$ENDIF} Dest: TSherpaOnnxOfflineModelConfig);
     function ToString: AnsiString;
   end;
@@ -694,6 +700,9 @@ type
   SherpaOnnxOfflineNemoEncDecCtcModelConfig = record
     Model: PAnsiChar;
   end;
+  SherpaOnnxOfflineDolphinModelConfig = record
+    Model: PAnsiChar;
+  end;
   SherpaOnnxOfflineWhisperModelConfig = record
     Encoder: PAnsiChar;
     Decoder: PAnsiChar;
@@ -740,6 +749,7 @@ type
     SenseVoice:  SherpaOnnxOfflineSenseVoiceModelConfig;
     Moonshine: SherpaOnnxOfflineMoonshineModelConfig;
     FireRedAsr: SherpaOnnxOfflineFireRedAsrModelConfig;
+    Dolphin: SherpaOnnxOfflineDolphinModelConfig;
   end;
 
   SherpaOnnxOfflineRecognizerConfig = record
@@ -1461,6 +1471,12 @@ begin
     [Self.Model]);
 end;
 
+function TSherpaOnnxOfflineDolphinModelConfig.ToString: AnsiString;
+begin
+  Result := Format('TSherpaOnnxOfflineDolphinModelConfig(Model := %s)',
+    [Self.Model]);
+end;
+
 function TSherpaOnnxOfflineWhisperModelConfig.ToString: AnsiString;
 begin
   Result := Format('TSherpaOnnxOfflineWhisperModelConfig(' +
@@ -1534,14 +1550,15 @@ begin
     'TeleSpeechCtc := %s, ' +
     'SenseVoice := %s, ' +
     'Moonshine := %s, ' +
-    'FireRedAsr := %s' +
+    'FireRedAsr := %s, ' +
+    'Dolphin := %s' +
     ')',
     [Self.Transducer.ToString, Self.Paraformer.ToString,
      Self.NeMoCtc.ToString, Self.Whisper.ToString, Self.Tdnn.ToString,
      Self.Tokens, Self.NumThreads, Self.Debug.ToString, Self.Provider,
      Self.ModelType, Self.ModelingUnit, Self.BpeVocab,
      Self.TeleSpeechCtc, Self.SenseVoice.ToString, Self.Moonshine.ToString,
-     Self.FireRedAsr.ToString
+     Self.FireRedAsr.ToString, Self.Dolphin.ToString
      ]);
 end;
 
@@ -1610,6 +1627,8 @@ begin
   C.ModelConfig.FireRedAsr.Encoder := PAnsiChar(Config.ModelConfig.FireRedAsr.Encoder);
   C.ModelConfig.FireRedAsr.Decoder := PAnsiChar(Config.ModelConfig.FireRedAsr.Decoder);
 
+  C.ModelConfig.Dolphin.Model := PAnsiChar(Config.ModelConfig.Dolphin.Model);
+
   C.LMConfig.Model := PAnsiChar(Config.LMConfig.Model);
   C.LMConfig.Scale := Config.LMConfig.Scale;