Add ten-vad Rust API example to remove silences from a file (#3778)

.github/scripts/test-rust.sh

@@ -81,6 +81,8 @@ rm -rf sherpa-onnx-punct-ct-transformer-zh-en-vocab272727-2024-04-12-int8
 
 ./run-silero-vad-remove-silence.sh
 
+./run-ten-vad-remove-silence.sh
+
 ./run-nemo-parakeet-en.sh
 ./run-zipformer-vi.sh
 ./run-zipformer-zh-en.sh

rust-api-examples/README.md

@@ -91,6 +91,7 @@ in your own Cargo project, see
 | 46 | [zipformer_transducer_simulate_streaming_microphone](#example-46-simulated-streaming-asr-with-zipformer-transducer-japanese-and-vad-from-microphone) | Simulated streaming ASR with Zipformer transducer (Japanese) and VAD from microphone |
 | 47 | [qwen3_asr_simulate_streaming_microphone](#example-47-simulated-streaming-asr-with-qwen3-asr-and-vad-from-microphone) | Simulated streaming ASR with Qwen3 ASR and VAD from microphone |
 | 48 | [whisper](#example-48-asr-with-non-streaming-whisper) | Non-streaming ASR with Whisper (multilingual) |
+| 49 | [ten_vad_remove_silence](#example-49-remove-silences-from-a-file-using-ten-vad) | Remove silences from an audio file using ten-vad |
 
 ## Run it
 
@@ -422,3 +423,9 @@ Qwen3 ASR recognizer on each detected segment.
 ```bash
 ./run-whisper.sh
 ```
+
+### Example 49: Remove silences from a file using ten-vad
+
+```bash
+./run-ten-vad-remove-silence.sh
+```

rust-api-examples/examples/ten_vad_remove_silence.rs

@@ -0,0 +1,122 @@
+// Copyright (c) 2026 Xiaomi Corporation
+//
+// This file demonstrates how to use ten-vad with sherpa-onnx's
+// Rust API to remove non-speech segments and save speech-only audio.
+//
+// See ../README.md for how to run it
+
+use clap::Parser;
+use sherpa_onnx::{self, TenVadModelConfig, VadModelConfig, VoiceActivityDetector, Wave};
+
+/// Simple VAD example: remove non-speech segments from a WAV file using ten-vad
+#[derive(Parser, Debug)]
+#[command(author, version, about, long_about = None)]
+struct Args {
+    /// Path to input WAV file
+    #[arg(long)]
+    input: String,
+
+    /// Path to output WAV file
+    #[arg(long)]
+    output: String,
+
+    /// Path to ten-vad ONNX model
+    #[arg(long)]
+    ten_vad_model: String,
+}
+
+fn main() -> anyhow::Result<()> {
+    let args = Args::parse();
+
+    // Read WAV file
+    let wave = Wave::read(&args.input)
+        .ok_or_else(|| anyhow::anyhow!("Failed to read WAV file: {}", &args.input))?;
+    let sample_rate = wave.sample_rate();
+    let input_num_samples = wave.num_samples();
+    let input_duration = input_num_samples as f32 / sample_rate as f32;
+
+    println!(
+        "Input WAV: sample rate: {}, num samples: {}, duration: {:.2}s",
+        sample_rate, input_num_samples, input_duration
+    );
+
+    if sample_rate != 16000 {
+        anyhow::bail!(
+            "ten-vad expects a sample rate of 16000 Hz, but the input file has {} Hz",
+            sample_rate
+        );
+    }
+
+    // Configure VAD
+    let mut ten_vad_config = TenVadModelConfig::default();
+    ten_vad_config.model = Some(args.ten_vad_model);
+
+    // ten-vad expects a window size of 256 samples at 16 kHz.
+    // Please don't change it unless you know what you are doing.
+    ten_vad_config.window_size = 256;
+
+    // You can tune the values below
+    ten_vad_config.threshold = 0.5;
+    ten_vad_config.min_silence_duration = 0.25;
+    ten_vad_config.min_speech_duration = 0.5;
+    ten_vad_config.max_speech_duration = 5.0;
+
+    let window_size = ten_vad_config.window_size as usize;
+
+    let vad_config = VadModelConfig {
+        silero_vad: Default::default(),
+        ten_vad: ten_vad_config,
+        sample_rate,
+        num_threads: 1,
+        provider: Some("cpu".to_string()),
+        debug: false,
+    };
+
+    let vad = VoiceActivityDetector::create(&vad_config, 30.0)
+        .ok_or_else(|| anyhow::anyhow!("Failed to create VoiceActivityDetector"))?;
+
+    let mut speech_samples = Vec::with_capacity(input_num_samples as usize);
+
+    for chunk in wave.samples().chunks(window_size) {
+        vad.accept_waveform(chunk);
+
+        while let Some(seg) = vad.front() {
+            speech_samples.extend_from_slice(seg.samples());
+            vad.pop();
+        }
+    }
+
+    vad.flush();
+    while let Some(seg) = vad.front() {
+        speech_samples.extend_from_slice(seg.samples());
+        vad.pop();
+    }
+
+    // Write speech-only samples to output WAV
+    let ok = sherpa_onnx::write(&args.output, &speech_samples, sample_rate);
+    if !ok {
+        anyhow::bail!("Failed to save speech-only audio to {}", args.output);
+    }
+    println!("Saved speech-only audio to {}", args.output);
+
+    // Summary
+    let output_num_samples = speech_samples.len();
+    let output_duration = output_num_samples as f32 / sample_rate as f32;
+    println!("\n=== Summary ===");
+    println!(
+        "Input:  sample rate = {}, samples = {}, duration = {:.2}s",
+        sample_rate, input_num_samples, input_duration
+    );
+    println!(
+        "Output: sample rate = {}, samples = {}, duration = {:.2}s",
+        sample_rate, output_num_samples, output_duration
+    );
+    if input_duration > 0.0 {
+        println!(
+            "Removed non-speech: {:.2}% of input removed",
+            100.0 * (1.0 - output_duration / input_duration)
+        );
+    }
+
+    Ok(())
+}

rust-api-examples/run-ten-vad-remove-silence.sh

@@ -0,0 +1,16 @@
+#!/usr/bin/env bash
+set -ex
+
+# https://k2-fsa.github.io/sherpa/onnx/vad/index.html
+if [ ! -f "./ten-vad.onnx" ]; then
+  curl -SLf -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/ten-vad.onnx
+fi
+
+if [ ! -f ./lei-jun-test.wav ]; then
+  curl -SLf -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/lei-jun-test.wav
+fi
+
+cargo run --example ten_vad_remove_silence -- \
+    --input ./lei-jun-test.wav \
+    --output ./no-silence-ten-vad.wav \
+    --ten-vad-model ./ten-vad.onnx