Add support for latest DPDFNet models and offline attenuation limit

c-api-examples/speech-enhancement-dpdfnet-c-api.c

@@ -17,13 +17,17 @@ wget https://github.com/k2-fsa/sherpa-onnx/releases/download/speech-enhancement-
 wget https://github.com/k2-fsa/sherpa-onnx/releases/download/speech-enhancement-models/dpdfnet4.onnx
 wget https://github.com/k2-fsa/sherpa-onnx/releases/download/speech-enhancement-models/dpdfnet8.onnx
 wget https://github.com/k2-fsa/sherpa-onnx/releases/download/speech-enhancement-models/dpdfnet2_48khz_hr.onnx
+wget https://huggingface.co/Ceva-IP/DPDFNet/resolve/main/onnx/dpdfnet2_8khz.onnx
+wget https://huggingface.co/Ceva-IP/DPDFNet/resolve/main/onnx/dpdfnet8_8khz.onnx
+wget https://huggingface.co/Ceva-IP/DPDFNet/resolve/main/onnx/dpdfnet8_48khz_hr.onnx
 wget https://github.com/k2-fsa/sherpa-onnx/releases/download/speech-enhancement-models/inp_16k.wav
 */
 // clang-format on
 //
 // Use dpdfnet_baseline.onnx, dpdfnet2.onnx, dpdfnet4.onnx, or dpdfnet8.onnx
 // for 16 kHz downstream ASR or speech recognition.
-// Use dpdfnet2_48khz_hr.onnx for 48 kHz enhancement output.
+// Use dpdfnet2_8khz.onnx or dpdfnet8_8khz.onnx for 8 kHz enhancement, and
+// dpdfnet2_48khz_hr.onnx or dpdfnet8_48khz_hr.onnx for 48 kHz enhancement.
 #include <stdio.h>
 #include <string.h>
 
@@ -37,6 +41,7 @@ int32_t main() {
 
   memset(&config, 0, sizeof(config));
   config.model.dpdfnet.model = model_filename;
+  config.model.dpdfnet.attenuation_limit_db = 12.0f;
 
   const SherpaOnnxOfflineSpeechDenoiser *sd =
       SherpaOnnxCreateOfflineSpeechDenoiser(&config);

cxx-api-examples/speech-enhancement-dpdfnet-cxx-api.cc

@@ -34,6 +34,7 @@ int32_t main() {
   std::string wav_filename = "./inp_16k.wav";
   std::string out_wave_filename = "./enhanced-dpdfnet.wav";
   config.model.dpdfnet.model = model_filename;
+  config.model.dpdfnet.attenuation_limit_db = 12.0f;
 
   auto sd = OfflineSpeechDenoiser::Create(config);
   if (!sd.Get()) {

dart-api-examples/speech-enhancement-dpdfnet/bin/speech_enhancement_dpdfnet.dart

@@ -26,13 +26,17 @@ void main(List<String> arguments) async {
   final outputWav = res['output-wav'] as String;
 
   final config = sherpa_onnx.OfflineSpeechDenoiserConfig(
-      model: sherpa_onnx.OfflineSpeechDenoiserModelConfig(
-    gtcrn: const sherpa_onnx.OfflineSpeechDenoiserGtcrnModelConfig(),
-    dpdfnet: sherpa_onnx.OfflineSpeechDenoiserDpdfNetModelConfig(model: model),
-    numThreads: 1,
-    debug: true,
-    provider: 'cpu',
-  ));
+    model: sherpa_onnx.OfflineSpeechDenoiserModelConfig(
+      gtcrn: const sherpa_onnx.OfflineSpeechDenoiserGtcrnModelConfig(),
+      dpdfnet: sherpa_onnx.OfflineSpeechDenoiserDpdfNetModelConfig(
+        model: model,
+        attenuationLimitDb: 12.0,
+      ),
+      numThreads: 1,
+      debug: true,
+      provider: 'cpu',
+    ),
+  );
 
   final sd = sherpa_onnx.OfflineSpeechDenoiser(config);
 

dotnet-examples/speech-enhancement-dpdfnet/Program.cs

@@ -31,6 +31,7 @@ static void Main(string[] args)
     var model = "./dpdfnet_baseline.onnx";
     var config = new OfflineSpeechDenoiserConfig();
     config.Model.Dpdfnet.Model = model;
+    config.Model.Dpdfnet.AttenuationLimitDb = 12.0f;
     config.Model.Debug = 1;
     config.Model.NumThreads = 1;
     var sd = new OfflineSpeechDenoiser(config);

flutter/sherpa_onnx/lib/src/offline_speech_denoiser.dart

@@ -23,6 +23,8 @@ class OfflineSpeechDenoiser {
     final c = calloc<SherpaOnnxOfflineSpeechDenoiserConfig>();
     c.ref.model.gtcrn.model = config.model.gtcrn.model.toNativeUtf8();
     c.ref.model.dpdfnet.model = config.model.dpdfnet.model.toNativeUtf8();
+    c.ref.model.dpdfnet.attenuationLimitDb =
+        config.model.dpdfnet.attenuationLimitDb;
 
     c.ref.model.numThreads = config.model.numThreads;
     c.ref.model.debug = config.model.debug ? 1 : 0;

flutter/sherpa_onnx/lib/src/offline_speech_denoiser_config.dart

@@ -35,25 +35,31 @@ class OfflineSpeechDenoiserGtcrnModelConfig {
 class OfflineSpeechDenoiserDpdfNetModelConfig {
   const OfflineSpeechDenoiserDpdfNetModelConfig({
     this.model = '',
+    this.attenuationLimitDb = 0.0,
   });
 
   factory OfflineSpeechDenoiserDpdfNetModelConfig.fromJson(
       Map<String, dynamic> json) {
     return OfflineSpeechDenoiserDpdfNetModelConfig(
       model: json['model'] as String? ?? '',
+      attenuationLimitDb:
+          (json['attenuation_limit_db'] as num?)?.toDouble() ?? 0.0,
     );
   }
 
   @override
   String toString() {
-    return 'OfflineSpeechDenoiserDpdfNetModelConfig(model: $model)';
+    return 'OfflineSpeechDenoiserDpdfNetModelConfig(model: $model, '
+        'attenuationLimitDb: $attenuationLimitDb)';
   }
 
   Map<String, dynamic> toJson() => {
         'model': model,
+        'attenuation_limit_db': attenuationLimitDb,
       };
 
   final String model;
+  final double attenuationLimitDb;
 }
 
 /// Aggregate model configuration for [OfflineSpeechDenoiser].

flutter/sherpa_onnx/lib/src/sherpa_onnx_bindings.dart

@@ -8,6 +8,9 @@ final class SherpaOnnxOfflineSpeechDenoiserGtcrnModelConfig extends Struct {
 
 final class SherpaOnnxOfflineSpeechDenoiserDpdfNetModelConfig extends Struct {
   external Pointer<Utf8> model;
+
+  @Float()
+  external double attenuationLimitDb;
 }
 
 final class SherpaOnnxOfflineSpeechDenoiserModelConfig extends Struct {

go-api-examples/speech-enhancement-dpdfnet/main.go

@@ -10,6 +10,7 @@ func main() {
 
 	config := sherpa.OfflineSpeechDenoiserConfig{}
 	config.Model.DpdfNet.Model = "./dpdfnet_baseline.onnx"
+	config.Model.DpdfNet.AttenuationLimitDb = 12
 	config.Model.NumThreads = 1
 	config.Model.Debug = 1
 

harmony-os/SherpaOnnxHar/sherpa_onnx/src/main/cpp/speech-denoiser.h

@@ -36,6 +36,7 @@ GetSpeechDenoiserDpdfNetModelConfig(Napi::Object obj) {
 
   Napi::Object o = obj.Get("dpdfnet").As<Napi::Object>();
   SHERPA_ONNX_ASSIGN_ATTR_STR(model, model);
+  SHERPA_ONNX_ASSIGN_ATTR_FLOAT(attenuation_limit_db, attenuationLimitDb);
   return c;
 }
 

java-api-examples/NonStreamingSpeechEnhancementDpdfNet.java

@@ -21,11 +21,16 @@ public static void main(String[] args) {
             .setDebug(true)
             .setProvider("cpu")
             .setDpdfnet(
-                OfflineSpeechDenoiserDpdfNetModelConfig.builder().setModel(model).build());
+                OfflineSpeechDenoiserDpdfNetModelConfig.builder()
+                    .setModel(model)
+                    .setAttenuationLimitDb(12.0f)
+                    .build());
 
     OfflineSpeechDenoiserModelConfig modelConfig = builder.build();
     OfflineSpeechDenoiserConfig config =
-        OfflineSpeechDenoiserConfig.builder().setModel(modelConfig).build();
+        OfflineSpeechDenoiserConfig.builder()
+            .setModel(modelConfig)
+            .build();
 
     OfflineSpeechDenoiser speech_denoiser = new OfflineSpeechDenoiser(config);
 

kotlin-api-examples/test_offline_speech_denoiser_dpdfnet.kt

@@ -14,7 +14,8 @@ fun createOfflineSpeechDenoiserDpdfNet(): OfflineSpeechDenoiser {
   val config = OfflineSpeechDenoiserConfig(
       model = OfflineSpeechDenoiserModelConfig(
         dpdfnet = OfflineSpeechDenoiserDpdfNetModelConfig(
-          model = "./dpdfnet_baseline.onnx"
+          model = "./dpdfnet_baseline.onnx",
+          attenuationLimitDb = 12.0f,
         ),
         provider = "cpu",
         numThreads = 1,

nodejs-addon-examples/test_offline_speech_enhancement_dpdfnet.js

@@ -7,7 +7,10 @@ function createOfflineSpeechDenoiser() {
   // https://github.com/k2-fsa/sherpa-onnx/releases/tag/speech-enhancement-models
   const config = {
     model: {
-      dpdfnet: {model: './dpdfnet_baseline.onnx'},
+      dpdfnet: {
+        model: './dpdfnet_baseline.onnx',
+        attenuationLimitDb: 12,
+      },
       debug: true,
       numThreads: 1,
     },

nodejs-examples/test-offline-speech-enhancement-dpdfnet.js

@@ -15,7 +15,7 @@ function createOfflineSpeechDenoiser() {
   const model = './dpdfnet2.onnx';
   let config = {
     model: {
-      dpdfnet: {model},
+      dpdfnet: {model, attenuationLimitDb: 12},
       debug: 1,
     },
   };

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

@@ -30,6 +30,7 @@
 
   Initialize(Config);
   Config.Model.DpdfNet.Model := './dpdfnet_baseline.onnx';
+  Config.Model.DpdfNet.AttenuationLimitDb := 12.0;
   Config.Model.NumThreads:= 1;
   Config.Model.Debug:= True;
   Config.Model.Provider:= 'cpu';

python-api-examples/README.md

@@ -21,9 +21,10 @@
 
 - [online-speech-enhancement-dpdfnet.py](./online-speech-enhancement-dpdfnet.py)
   It shows how to use the online speech denoiser API with DPDFNet.
-  models. Use 16 kHz DPDFNet models such as `dpdfnet_baseline.onnx`,
+  Use 16 kHz DPDFNet models such as `dpdfnet_baseline.onnx`,
   `dpdfnet2.onnx`, `dpdfnet4.onnx`, or `dpdfnet8.onnx` for downstream ASR and
-  `dpdfnet2_48khz_hr.onnx` for 48 kHz enhancement output.
+  `dpdfnet2_8khz.onnx` or `dpdfnet8_8khz.onnx` for 8 kHz enhancement.
+  Use `dpdfnet2_48khz_hr.onnx` or `dpdfnet8_48khz_hr.onnx` for 48 kHz output.
 
 - [pocket-tts.py](./pocket-tts.py) It shows how to use PocketTTS with the
   `GenerationConfig` API.

python-api-examples/offline-speech-enhancement-dpdfnet.py

@@ -15,11 +15,15 @@
  wget https://github.com/k2-fsa/sherpa-onnx/releases/download/speech-enhancement-models/dpdfnet4.onnx
  wget https://github.com/k2-fsa/sherpa-onnx/releases/download/speech-enhancement-models/dpdfnet8.onnx
  wget https://github.com/k2-fsa/sherpa-onnx/releases/download/speech-enhancement-models/dpdfnet2_48khz_hr.onnx
+ wget https://huggingface.co/Ceva-IP/DPDFNet/resolve/main/onnx/dpdfnet2_8khz.onnx
+ wget https://huggingface.co/Ceva-IP/DPDFNet/resolve/main/onnx/dpdfnet8_8khz.onnx
+ wget https://huggingface.co/Ceva-IP/DPDFNet/resolve/main/onnx/dpdfnet8_48khz_hr.onnx
  wget https://github.com/k2-fsa/sherpa-onnx/releases/download/speech-enhancement-models/speech_with_noise.wav
 
 Use 16 kHz DPDFNet models such as `dpdfnet_baseline.onnx`, `dpdfnet2.onnx`,
 `dpdfnet4.onnx`, or `dpdfnet8.onnx` for downstream ASR or speech recognition.
-Use `dpdfnet2_48khz_hr.onnx` for 48 kHz enhancement output.
+Use `dpdfnet2_8khz.onnx` or `dpdfnet8_8khz.onnx` for 8 kHz enhancement, and
+`dpdfnet2_48khz_hr.onnx` or `dpdfnet8_48khz_hr.onnx` for 48 kHz output.
 """
 
 import time
@@ -45,12 +49,13 @@ def create_speech_denoiser():
     config = sherpa_onnx.OfflineSpeechDenoiserConfig(
         model=sherpa_onnx.OfflineSpeechDenoiserModelConfig(
             dpdfnet=sherpa_onnx.OfflineSpeechDenoiserDpdfNetModelConfig(
-                model=model_filename
+                model=model_filename,
+                attenuation_limit_db=12.0,
             ),
             debug=False,
             num_threads=1,
             provider="cpu",
-        )
+        ),
     )
     if not config.validate():
         print(config)

rust-api-examples/examples/offline_speech_enhancement_dpdfnet.rs

@@ -24,6 +24,7 @@ fn main() -> anyhow::Result<()> {
         model: sherpa_onnx::OfflineSpeechDenoiserModelConfig {
             dpdfnet: OfflineSpeechDenoiserDpdfNetModelConfig {
                 model: Some(args.model),
+                attenuation_limit_db: 12.0,
             },
             ..Default::default()
         },

rust-api-examples/examples/streaming_speech_enhancement_dpdfnet.rs

@@ -1,7 +1,7 @@
 use clap::Parser;
 use sherpa_onnx::{
-    write, OfflineSpeechDenoiserDpdfNetModelConfig, OnlineSpeechDenoiser, OnlineSpeechDenoiserConfig,
-    Wave,
+    write, OfflineSpeechDenoiserDpdfNetModelConfig, OnlineSpeechDenoiser,
+    OnlineSpeechDenoiserConfig, Wave,
 };
 
 #[derive(Parser, Debug)]
@@ -24,6 +24,7 @@ fn main() -> anyhow::Result<()> {
         model: sherpa_onnx::OfflineSpeechDenoiserModelConfig {
             dpdfnet: OfflineSpeechDenoiserDpdfNetModelConfig {
                 model: Some(args.model),
+                ..Default::default()
             },
             ..Default::default()
         },

scripts/dotnet/OfflineSpeechDenoiserDpdfNetModelConfig.cs

@@ -10,9 +10,11 @@ public struct OfflineSpeechDenoiserDpdfNetModelConfig
         public OfflineSpeechDenoiserDpdfNetModelConfig()
         {
             Model = "";
+            AttenuationLimitDb = 0.0f;
         }
 
         [MarshalAs(UnmanagedType.LPStr)]
         public string Model;
+        public float AttenuationLimitDb;
     }
 }

scripts/go/sherpa_onnx.go

@@ -2570,7 +2570,8 @@ type OfflineSpeechDenoiserGtcrnModelConfig struct {
 }
 
 type OfflineSpeechDenoiserDpdfNetModelConfig struct {
-	Model string
+	Model              string
+	AttenuationLimitDb float32
 }
 
 type OfflineSpeechDenoiserModelConfig struct {
@@ -2650,6 +2651,7 @@ func NewOfflineSpeechDenoiser(config *OfflineSpeechDenoiserConfig) *OfflineSpeec
 	defer C.free(unsafe.Pointer(c.model.gtcrn.model))
 	c.model.dpdfnet.model = C.CString(config.Model.DpdfNet.Model)
 	defer C.free(unsafe.Pointer(c.model.dpdfnet.model))
+	c.model.dpdfnet.attenuation_limit_db = C.float(config.Model.DpdfNet.AttenuationLimitDb)
 
 	c.model.num_threads = C.int(config.Model.NumThreads)
 	c.model.debug = C.int(config.Model.Debug)

scripts/node-addon-api/lib/types.js

@@ -616,9 +616,10 @@
  */
 
 /**
- * Offline Speech Denoiser model config
- * @typedef {Object} OfflineSpeechDenoiserDpdfNetModelConfig
- * @property {string} [model]
+ * Offline Speech Denoiser model config
+ * @typedef {Object} OfflineSpeechDenoiserDpdfNetModelConfig
+ * @property {string} [model]
+ * @property {number} [attenuationLimitDb] Offline attenuation limit in dB.
  */
 
 /**
@@ -632,10 +633,10 @@
  */
 
 /**
- * Offline Speech Denoiser configuration (partial).
- * @typedef {Object} OfflineSpeechDenoiserConfig
- * @property {OfflineSpeechDenoiserModelConfig} [model]
- */
+ * Offline Speech Denoiser configuration (partial).
+ * @typedef {Object} OfflineSpeechDenoiserConfig
+ * @property {OfflineSpeechDenoiserModelConfig} [model]
+ */
 
 /**
  * Online Speech Denoiser configuration (partial).

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

@@ -2828,6 +2828,8 @@ static sherpa_onnx::OfflineSpeechDenoiserConfig GetOfflineSpeechDenoiserConfig(
   c.model.debug = config->model.debug;
   c.model.provider = SHERPA_ONNX_OR(config->model.provider, "cpu");
   c.model.dpdfnet.model = SHERPA_ONNX_OR(config->model.dpdfnet.model, "");
+  c.model.dpdfnet.attenuation_limit_db =
+      config->model.dpdfnet.attenuation_limit_db;
 
   if (c.model.debug) {
 #if __OHOS__

sherpa-onnx/c-api/c-api.h

@@ -4098,6 +4098,15 @@ typedef struct SherpaOnnxOfflineSpeechDenoiserGtcrnModelConfig {
 typedef struct SherpaOnnxOfflineSpeechDenoiserDpdfNetModelConfig {
   /** Model filename. */
   const char *model;
+  /**
+   * Offline attenuation limit in dB.
+   *
+   * Values greater than 0 limit suppression by blending aligned noisy spectra
+   * into the enhanced spectra with a noisy-signal weight of
+   * `10^(-limit_db / 20)`. The maximum finite value is 100. Set to 0 or
+   * infinity to disable the limit.
+   */
+  float attenuation_limit_db;
 } SherpaOnnxOfflineSpeechDenoiserDpdfNetModelConfig;
 
 /**

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

@@ -20,6 +20,7 @@ static void FillSpeechDenoiserModelConfig(
   memset(dst, 0, sizeof(*dst));
   dst->gtcrn.model = src.gtcrn.model.c_str();
   dst->dpdfnet.model = src.dpdfnet.model.c_str();
+  dst->dpdfnet.attenuation_limit_db = src.dpdfnet.attenuation_limit_db;
   dst->num_threads = src.num_threads;
   dst->provider = src.provider.c_str();
   dst->debug = src.debug;

sherpa-onnx/c-api/cxx-api.h

@@ -1219,6 +1219,8 @@ struct OfflineSpeechDenoiserGtcrnModelConfig {
 struct OfflineSpeechDenoiserDpdfNetModelConfig {
   /** Model ONNX file. */
   std::string model;
+  /** Offline attenuation limit in dB. 0 disables it. */
+  float attenuation_limit_db = 0.0f;
 };
 
 /**

sherpa-onnx/c-api/docs/speech-enhancement.dox

@@ -39,6 +39,7 @@ Example source: [speech-enhancement-gtcrn-c-api.c](https://github.com/k2-fsa/she
 SherpaOnnxOfflineSpeechDenoiserConfig config;
 memset(&config, 0, sizeof(config));
 config.model.dpdfnet.model = "./dpdfnet_baseline.onnx";
+config.model.dpdfnet.attenuation_limit_db = 12.0f;
 config.model.num_threads = 1;
 config.model.provider = "cpu";
 
@@ -49,7 +50,8 @@ SherpaOnnxDestroyOfflineSpeechDenoiser(sd);
 @endcode
 
 Model file: `dpdfnet_baseline.onnx` (also: `dpdfnet2.onnx`, `dpdfnet4.onnx`,
-`dpdfnet8.onnx` for 16 kHz; `dpdfnet2_48khz_hr.onnx` for 48 kHz)
+`dpdfnet8.onnx` for 16 kHz; `dpdfnet2_8khz.onnx`, `dpdfnet8_8khz.onnx` for
+8 kHz; and `dpdfnet2_48khz_hr.onnx`, `dpdfnet8_48khz_hr.onnx` for 48 kHz)
 
 Example source: [speech-enhancement-dpdfnet-c-api.c](https://github.com/k2-fsa/sherpa-onnx/blob/master/c-api-examples/speech-enhancement-dpdfnet-c-api.c)
 
@@ -92,7 +94,9 @@ const SherpaOnnxOnlineSpeechDenoiser *sd =
 SherpaOnnxDestroyOnlineSpeechDenoiser(sd);
 @endcode
 
-Model file: `dpdfnet_baseline.onnx`
+Model files: `dpdfnet_baseline.onnx`, `dpdfnet2.onnx`, `dpdfnet4.onnx`,
+`dpdfnet8.onnx` for 16 kHz; `dpdfnet2_8khz.onnx`, `dpdfnet8_8khz.onnx` for
+8 kHz; and `dpdfnet2_48khz_hr.onnx`, `dpdfnet8_48khz_hr.onnx` for 48 kHz.
 
 Example source: [online-speech-enhancement-dpdfnet-c-api.c](https://github.com/k2-fsa/sherpa-onnx/blob/master/c-api-examples/online-speech-enhancement-dpdfnet-c-api.c)
 */