Add Pascal API for speech enhancement GTCRN models (#1992)

.github/workflows/pascal.yaml

@@ -111,20 +111,36 @@ jobs:
           ls -lh install/lib/
 
           if [[ ${{ matrix.os }} == 'windows-latest' ]]; then
-            cp -v install/lib/*.dll ../pascal-api-examples/read-wav
-            cp -v install/lib/*.dll ../pascal-api-examples/streaming-asr
             cp -v install/lib/*.dll ../pascal-api-examples/non-streaming-asr
+            cp -v install/lib/*.dll ../pascal-api-examples/read-wav
+            cp -v install/lib/*.dll ../pascal-api-examples/speaker-diarization
+            cp -v install/lib/*.dll ../pascal-api-examples/speech-enhancement-gtcrn
+            cp -v install/lib/*.dll ../pascal-api-examples/streaming-asr
+            cp -v install/lib/*.dll ../pascal-api-examples/tts
             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/*.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/read-wav
+            cp -v ../sherpa-onnx/pascal-api/*.pas ../pascal-api-examples/speaker-diarization
+            cp -v ../sherpa-onnx/pascal-api/*.pas ../pascal-api-examples/speech-enhancement-gtcrn
+            cp -v ../sherpa-onnx/pascal-api/*.pas ../pascal-api-examples/streaming-asr
+            cp -v ../sherpa-onnx/pascal-api/*.pas ../pascal-api-examples/tts
             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 Speech Enhancement test (GTCRN)
+        shell: bash
+        run: |
+          export PATH=/c/lazarus/fpc/3.2.2/bin/x86_64-win64:$PATH
+
+          cd ./pascal-api-examples
+
+          pushd speech-enhancement-gtcrn
+          ./run-gtcrn.sh
+          ls -lh
+
       - name:  Run Pascal test (Non Streaming ASR)
         shell: bash
         run: |

.github/workflows/test-go.yaml

@@ -99,11 +99,21 @@ jobs:
             cp -v ./install/lib/sherpa-onnx-c-api.dll ../scripts/go/_internal/lib/x86_64-pc-windows-gnu/
             cp -v ./install/lib/onnxruntime.dll ../scripts/go/_internal/lib/x86_64-pc-windows-gnu/
             ls -lh ../scripts/go/_internal/lib/x86_64-pc-windows-gnu/
-            cp -v ../scripts/go/_internal/lib/x86_64-pc-windows-gnu/*.dll ../scripts/go/_internal/speaker-identification/
+            cp -v ../scripts/go/_internal/lib/x86_64-pc-windows-gnu/*.dll ../scripts/go/_internal/add-punctuation
-            cp -v ../scripts/go/_internal/lib/x86_64-pc-windows-gnu/*.dll ../scripts/go/_internal/streaming-hlg-decoding/
+            cp -v ../scripts/go/_internal/lib/x86_64-pc-windows-gnu/*.dll ../scripts/go/_internal/audio-tagging
-            cp -v ../scripts/go/_internal/lib/x86_64-pc-windows-gnu/*.dll ../scripts/go/_internal/non-streaming-tts/
+            cp -v ../scripts/go/_internal/lib/x86_64-pc-windows-gnu/*.dll ../scripts/go/_internal/keyword-spotting-from-file/
             cp -v ../scripts/go/_internal/lib/x86_64-pc-windows-gnu/*.dll ../scripts/go/_internal/non-streaming-decode-files/
+            cp -v ../scripts/go/_internal/lib/x86_64-pc-windows-gnu/*.dll ../scripts/go/_internal/non-streaming-speaker-diarization/
+            cp -v ../scripts/go/_internal/lib/x86_64-pc-windows-gnu/*.dll ../scripts/go/_internal/non-streaming-tts/
+            cp -v ../scripts/go/_internal/lib/x86_64-pc-windows-gnu/*.dll ../scripts/go/_internal/speaker-identification/
+            cp -v ../scripts/go/_internal/lib/x86_64-pc-windows-gnu/*.dll ../scripts/go/_internal/speech-enhancement-gtcrn
             cp -v ../scripts/go/_internal/lib/x86_64-pc-windows-gnu/*.dll ../scripts/go/_internal/streaming-decode-files/
+            cp -v ../scripts/go/_internal/lib/x86_64-pc-windows-gnu/*.dll ../scripts/go/_internal/streaming-hlg-decoding/
+            cp -v ../scripts/go/_internal/lib/x86_64-pc-windows-gnu/*.dll ../scripts/go/_internal/vad
+            cp -v ../scripts/go/_internal/lib/x86_64-pc-windows-gnu/*.dll ../scripts/go/_internal/vad-asr-paraformer
+            cp -v ../scripts/go/_internal/lib/x86_64-pc-windows-gnu/*.dll ../scripts/go/_internal/vad-asr-whisper
+            cp -v ../scripts/go/_internal/lib/x86_64-pc-windows-gnu/*.dll ../scripts/go/_internal/vad-speaker-identification
+            cp -v ../scripts/go/_internal/lib/x86_64-pc-windows-gnu/*.dll ../scripts/go/_internal/vad-spoken-language-identification
 
             cp -v ../scripts/go/_internal/lib/x86_64-pc-windows-gnu/*.dll $upload_dir
           else

pascal-api-examples/speech-enhancement-gtcrn/.gitignore

@@ -0,0 +1 @@
+gtcrn

pascal-api-examples/speech-enhancement-gtcrn/gtcrn.pas

@@ -0,0 +1,43 @@
+{ Copyright (c)  2025  Xiaomi Corporation }
+{
+This file shows how to use the speech enhancement API from sherpa-onnx
+
+Please first download files used in this script before you run it.
+
+wget https://github.com/k2-fsa/sherpa-onnx/releases/download/speech-enhancement-models/gtcrn_simple.onnx
+wget https://github.com/k2-fsa/sherpa-onnx/releases/download/speech-enhancement-models/inp_16k.wav
+}
+program main;
+
+{$mode delphi}
+
+uses
+  sherpa_onnx,
+  SysUtils;
+
+var
+  Wave: TSherpaOnnxWave;
+
+  Config: TSherpaOnnxOfflineSpeechDenoiserConfig;
+  Sd: TSherpaOnnxOfflineSpeechDenoiser;
+  Audio: TSherpaOnnxDenoisedAudio;
+begin
+  Wave := SherpaOnnxReadWave('./inp_16k.wav');
+
+  Initialize(Config);
+
+  Config.Model.Gtcrn.Model := './gtcrn_simple.onnx';
+  Config.Model.NumThreads:= 1;
+  Config.Model.Debug:= True;
+  Config.Model.Provider:= 'cpu';
+
+  Sd := TSherpaOnnxOfflineSpeechDenoiser.Create(Config);
+
+  Audio := Sd.Run(Wave.Samples, Wave.SampleRate);
+
+  SherpaOnnxWriteWave('./enhanced-16k.wav', Audio.Samples, Audio.SampleRate);
+  WriteLn('Saved to ./enhanced-16k.wav');
+
+  FreeAndNil(Sd);
+end.
+

pascal-api-examples/speech-enhancement-gtcrn/run-gtcrn.sh

@@ -0,0 +1,44 @@
+#!/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: $SHERPA_ONNX_DIR"
+
+if [[ ! -f ../../build/install/lib/libsherpa-onnx-c-api.dylib  && ! -f ../../build/install/lib/libsherpa-onnx-c-api.so && ! -f ../../build/install/lib/sherpa-onnx-c-api.dll ]]; then
+  mkdir -p ../../build
+  pushd ../../build
+  cmake \
+    -DCMAKE_INSTALL_PREFIX=./install \
+    -DSHERPA_ONNX_ENABLE_PYTHON=OFF \
+    -DSHERPA_ONNX_ENABLE_TESTS=OFF \
+    -DSHERPA_ONNX_ENABLE_CHECK=OFF \
+    -DBUILD_SHARED_LIBS=ON \
+    -DSHERPA_ONNX_ENABLE_PORTAUDIO=OFF \
+    ..
+
+  cmake --build . --target install --config Release
+  popd
+fi
+
+if [ ! -f ./gtcrn_simple.onnx ]; then
+  curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/speech-enhancement-models/gtcrn_simple.onnx
+fi
+
+if [ ! -f ./inp_16k.wav ]; then
+  curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/speech-enhancement-models/inp_16k.wav
+fi
+
+fpc \
+  -dSHERPA_ONNX_USE_SHARED_LIBS \
+  -Fu$SHERPA_ONNX_DIR/sherpa-onnx/pascal-api \
+  -Fl$SHERPA_ONNX_DIR/build/install/lib \
+  ./gtcrn.pas
+
+export LD_LIBRARY_PATH=$SHERPA_ONNX_DIR/build/install/lib:$LD_LIBRARY_PATH
+export DYLD_LIBRARY_PATH=$SHERPA_ONNX_DIR/build/install/lib:$DYLD_LIBRARY_PATH
+
+./gtcrn
+

sherpa-onnx/pascal-api/sherpa_onnx.pas

@@ -515,6 +515,44 @@ type
     property GetSampleRate: Integer Read SampleRate;
   end;
 
+  TSherpaOnnxOfflineSpeechDenoiserGtcrnModelConfig = record
+    Model: AnsiString;
+    function ToString: AnsiString;
+  end;
+
+  TSherpaOnnxOfflineSpeechDenoiserModelConfig = record
+    Gtcrn: TSherpaOnnxOfflineSpeechDenoiserGtcrnModelConfig;
+    NumThreads: Integer;
+    Debug: Boolean;
+    Provider: AnsiString;
+    function ToString: AnsiString;
+    class operator Initialize({$IFDEF FPC}var{$ELSE}out{$ENDIF} Dest: TSherpaOnnxOfflineSpeechDenoiserModelConfig);
+  end;
+
+  TSherpaOnnxOfflineSpeechDenoiserConfig = record
+    Model: TSherpaOnnxOfflineSpeechDenoiserModelConfig;
+    function ToString: AnsiString;
+  end;
+
+  TSherpaOnnxDenoisedAudio = record
+    Samples: array of Single;
+    SampleRate: Integer;
+  end;
+
+  TSherpaOnnxOfflineSpeechDenoiser = class
+  private
+   Handle: Pointer;
+   SampleRate: Integer;
+   _Config: TSherpaOnnxOfflineSpeechDenoiserConfig;
+  public
+    constructor Create(Config: TSherpaOnnxOfflineSpeechDenoiserConfig);
+    destructor Destroy; override;
+
+    function Run(Samples: array of Single; InputSampleRate: Integer): TSherpaOnnxDenoisedAudio;
+
+    property GetHandle: Pointer Read Handle;
+    property GetSampleRate: Integer Read SampleRate;
+  end;
 
   { It supports reading a single channel wave with 16-bit encoded samples.
     Samples are normalized to the range [-1, 1].
@@ -851,6 +889,31 @@ type
 
   PSherpaOnnxOfflineSpeakerDiarizationConfig = ^SherpaOnnxOfflineSpeakerDiarizationConfig;
 
+  SherpaOnnxOfflineSpeechDenoiserGtcrnModelConfig = record
+    Model: PAnsiChar;
+  end;
+
+  SherpaOnnxOfflineSpeechDenoiserModelConfig = record
+    Gtcrn: SherpaOnnxOfflineSpeechDenoiserGtcrnModelConfig;
+    NumThreads: cint32;
+    Debug: cint32;
+    Provider: PAnsiChar;
+  end;
+
+  SherpaOnnxOfflineSpeechDenoiserConfig = record
+    Model: SherpaOnnxOfflineSpeechDenoiserModelConfig;
+  end;
+
+  PSherpaOnnxOfflineSpeechDenoiserConfig = ^SherpaOnnxOfflineSpeechDenoiserConfig;
+
+  SherpaOnnxDenoisedAudio = record
+    Samples: pcfloat;
+    N: cint32;
+    SampleRate: cint32;
+  end;
+
+  PSherpaOnnxDenoisedAudio = ^SherpaOnnxDenoisedAudio;
+
 function SherpaOnnxCreateLinearResampler(SampleRateInHz: cint32;
   SampleRateOutHz: cint32;
   FilterCutoffHz: cfloat;
@@ -872,6 +935,22 @@ procedure SherpaOnnxLinearResamplerResampleFree(P: PSherpaOnnxResampleOut); cdec
 procedure SherpaOnnxLinearResamplerReset(P: Pointer); cdecl;
   external SherpaOnnxLibName;
 
+function SherpaOnnxCreateOfflineSpeechDenoiser(Config: PSherpaOnnxOfflineSpeechDenoiserConfig): Pointer; cdecl;
+  external SherpaOnnxLibName;
+
+procedure SherpaOnnxDestroyOfflineSpeechDenoiser(P: Pointer); cdecl;
+  external SherpaOnnxLibName;
+
+function SherpaOnnxOfflineSpeechDenoiserGetSampleRate(P: Pointer): cint32; cdecl;
+  external SherpaOnnxLibName;
+
+function SherpaOnnxOfflineSpeechDenoiserRun(P: Pointer;
+  Samples: pcfloat; N: cint32;SampleRate: cint32):PSherpaOnnxDenoisedAudio; cdecl;
+  external SherpaOnnxLibName;
+
+procedure SherpaOnnxDestroyDenoisedAudio(Audio: Pointer); cdecl;
+  external SherpaOnnxLibName;
+
 function SherpaOnnxCreateOfflineSpeakerDiarization(Config: PSherpaOnnxOfflineSpeakerDiarizationConfig): Pointer; cdecl;
   external SherpaOnnxLibName;
 
@@ -2358,4 +2437,79 @@ begin
   SherpaOnnxOfflineSpeakerDiarizationDestroyResult(R);
 end;
 
+function TSherpaOnnxOfflineSpeechDenoiserGtcrnModelConfig.ToString: AnsiString;
+begin
+  Result := Format('TSherpaOnnxOfflineSpeechDenoiserGtcrnModelConfig(' +
+    'Model := %s)', [Self.Model]);
+end;
+
+function TSherpaOnnxOfflineSpeechDenoiserModelConfig.ToString: AnsiString;
+begin
+  Result := Format('TSherpaOnnxOfflineSpeechDenoiserModelConfig(' +
+    'Gtcrn := %s, '+
+    'NumThreads := %d, '+
+    'Debug := %s, '+
+    'Provider := %s)',
+    [Self.Gtcrn.ToString, Self.NumThreads, Self.Debug.ToString, Self.Provider]);
+end;
+
+class operator TSherpaOnnxOfflineSpeechDenoiserModelConfig.Initialize({$IFDEF FPC}var{$ELSE}out{$ENDIF} Dest: TSherpaOnnxOfflineSpeechDenoiserModelConfig);
+begin
+  Dest.NumThreads := 1;
+  Dest.Debug := False;
+  Dest.Provider := 'cpu';
+end;
+
+function TSherpaOnnxOfflineSpeechDenoiserConfig.ToString: AnsiString;
+begin
+  Result := Format('TSherpaOnnxOfflineSpeechDenoiserConfig(' +
+    'Model := %s)', [Self.Model.ToString]);
+end;
+
+constructor TSherpaOnnxOfflineSpeechDenoiser.Create(Config: TSherpaOnnxOfflineSpeechDenoiserConfig);
+var
+  C: SherpaOnnxOfflineSpeechDenoiserConfig;
+begin
+  C := Default(SherpaOnnxOfflineSpeechDenoiserConfig);
+  C.Model.Gtcrn.Model := PAnsiChar(Config.Model.Gtcrn.Model);
+  C.Model.NumThreads := Config.Model.NumThreads;
+  C.Model.Debug := Ord(Config.Model.Debug);
+  C.Model.Provider := PAnsiChar(Config.Model.Provider);
+
+  Self.Handle := SherpaOnnxCreateOfflineSpeechDenoiser(@C);
+  Self._Config := Config;
+  Self.SampleRate :=  0;
+
+  if Self.Handle <> nil then
+    begin
+      Self.SampleRate := SherpaOnnxOfflineSpeechDenoiserGetSampleRate(Self.Handle);
+    end;
+end;
+
+destructor TSherpaOnnxOfflineSpeechDenoiser.Destroy;
+begin
+  SherpaOnnxDestroyOfflineSpeechDenoiser(Self.Handle);
+  Self.Handle := nil;
+end;
+
+function TSherpaOnnxOfflineSpeechDenoiser.Run(Samples: array of Single; InputSampleRate: Integer): TSherpaOnnxDenoisedAudio;
+var
+  Audio: PSherpaOnnxDenoisedAudio;
+  I: Integer;
+begin
+  Result := Default(TSherpaOnnxDenoisedAudio);
+
+  Audio := SherpaOnnxOfflineSpeechDenoiserRun(Self.Handle, pcfloat(Samples), Length(Samples), InputSampleRate);
+
+  SetLength(Result.Samples, Audio^.N);
+  Result.SampleRate := Audio^.SampleRate;
+
+  for I := Low(Result.Samples) to High(Result.Samples) do
+  begin
+    Result.Samples[I] := Audio^.Samples[I];
+  end;
+
+  SherpaOnnxDestroyDenoisedAudio(audio);
+end;
+
 end.