Add Silero VAD speech detection example

This exercises frequent use of the `If` operator, 1D convolutions and reflect
padding.

rten-examples/Cargo.toml

@@ -90,6 +90,10 @@ path = "src/qwen2_chat.rs"
 name = "piper"
 path = "src/piper.rs"
 
+[[bin]]
+name = "silero"
+path = "src/silero.rs"
+
 [[bin]]
 name = "wav2vec2"
 path = "src/wav2vec2.rs"

rten-examples/README.md

@@ -73,4 +73,5 @@ language model.
 ### Audio
 
 - **piper** - Text-to-speech using [Piper](https://github.com/rhasspy/piper) models
+- **silero** - Speech detection using [Silero VAD](https://github.com/snakers4/silero-vad)
 - **wav2vec2** - Speech recognition of .wav audio using [wav2vec2](https://arxiv.org/abs/2006.11477)

rten-examples/src/silero.rs

@@ -0,0 +1,291 @@
+use std::collections::VecDeque;
+use std::error::Error;
+use std::ops::Range;
+
+use rten::Model;
+use rten_tensor::prelude::*;
+use rten_tensor::NdTensor;
+
+struct Args {
+    model: String,
+    wav_file: String,
+}
+
+fn parse_args() -> Result<Args, lexopt::Error> {
+    use lexopt::prelude::*;
+
+    let mut values = VecDeque::new();
+
+    let mut parser = lexopt::Parser::from_env();
+    while let Some(arg) = parser.next()? {
+        match arg {
+            Value(val) => values.push_back(val.string()?),
+            Long("help") => {
+                println!(
+                    "Detect speech in .wav files.
+
+Usage: {bin_name} <model_path> <wav_file>
+",
+                    bin_name = parser.bin_name().unwrap_or("silero")
+                );
+                std::process::exit(0);
+            }
+            _ => return Err(arg.unexpected()),
+        }
+    }
+
+    let model = values.pop_front().ok_or("missing `model` arg")?;
+    let wav_file = values.pop_front().ok_or("missing `wav_file` arg")?;
+
+    let args = Args { model, wav_file };
+
+    Ok(args)
+}
+
+/// Read a .wav audio file into a sequence of samples with values in [1, -1].
+///
+/// The Silero model supports a sample rate of either 8 or 16 kHz. This example
+/// only supports 16 kHz.
+fn read_wav_file(path: &str) -> Result<Vec<f32>, hound::Error> {
+    let mut reader = hound::WavReader::open(path)?;
+
+    let spec = reader.spec();
+    let expected_sample_rate = 16_000;
+
+    if spec.sample_rate != expected_sample_rate {
+        println!(
+            "WARNING: Sample rate is {} kHz, this model expects {} kHz.",
+            spec.sample_rate / 1_000,
+            expected_sample_rate
+        );
+    }
+
+    let mut samples = Vec::new();
+    for sample in reader.samples::<i16>() {
+        samples.push(sample?);
+    }
+    let float_samples: Vec<f32> = samples
+        .into_iter()
+        .map(|x| (x as f32) / i16::MAX as f32)
+        .collect();
+    Ok(float_samples)
+}
+
+/// Time offset or duration in milliseconds.
+type Millis = usize;
+
+/// Configuration for the speech detection state machine.
+struct VadConfig {
+    /// Probability threshold indicating presence of speech
+    pos_threshold: f32,
+
+    /// Probability threshold indicating absence of speech
+    neg_threshold: f32,
+
+    /// Minimum duration of speech segments
+    min_speech_duration: Millis,
+
+    /// Minimum duration of a gap between two speech segments
+    min_silence_duration: Millis,
+}
+
+impl Default for VadConfig {
+    fn default() -> VadConfig {
+        // These are based on the defaults for the `get_speech_timestamps` function
+        // at https://github.com/snakers4/silero-vad/blob/46f94b7d6029e19b482eebdfff0c18012fa84675/src/silero_vad/utils_vad.py#L187.
+        VadConfig {
+            pos_threshold: 0.5,
+            neg_threshold: 0.35,
+            min_speech_duration: 250,
+            min_silence_duration: 100,
+        }
+    }
+}
+
+/// State machine which tracks whether voice activity is currently detected.
+struct VadState {
+    config: VadConfig,
+
+    /// True if we're currently in a speaking state.
+    speaking: bool,
+
+    /// Offset of start of chunk where last transition between speaking and
+    /// non-speaking occurred.
+    last_transition_time: Millis,
+
+    /// True if speech was detected in the previous audio chunk.
+    prev_chunk_speaking: bool,
+
+    /// End offset of previous audio chunk.
+    prev_time: Millis,
+}
+
+impl VadState {
+    fn new(config: VadConfig) -> VadState {
+        VadState {
+            config,
+            speaking: false,
+            last_transition_time: 0,
+            prev_chunk_speaking: false,
+            prev_time: 0,
+        }
+    }
+
+    /// Update the speech state using the speech probability output for an
+    /// audio chunk from the Silero model.
+    fn update(&mut self, current_time: Millis, speech_prob: f32) {
+        assert!(current_time >= self.prev_time);
+
+        let curr_chunk_speaking = if speech_prob > self.config.pos_threshold {
+            true
+        } else if speech_prob < self.config.neg_threshold {
+            false
+        } else {
+            self.prev_chunk_speaking
+        };
+        if curr_chunk_speaking != self.prev_chunk_speaking {
+            self.last_transition_time = self.prev_time;
+        }
+
+        if curr_chunk_speaking {
+            if !self.speaking {
+                let speech_duration = current_time - self.last_transition_time;
+                if speech_duration >= self.config.min_speech_duration {
+                    self.speaking = true;
+                }
+            }
+        } else if self.speaking {
+            let silence_duration = current_time - self.last_transition_time;
+            if silence_duration >= self.config.min_silence_duration {
+                self.speaking = false;
+            }
+        }
+
+        self.prev_chunk_speaking = curr_chunk_speaking;
+        self.prev_time = current_time;
+    }
+
+    /// Return the speech start time if speech was detected.
+    fn speech_start(&self) -> Option<Millis> {
+        self.speaking.then_some(self.last_transition_time)
+    }
+}
+
+/// Detect speech in .wav audio files using Silero VAD [^1].
+///
+/// Download the ONNX model from the Silero VAD repository at
+/// https://github.com/snakers4/silero-vad/tree/master/src/silero_vad/data,
+/// then convert it using:
+///
+/// ```
+/// rten-convert silero_vad.onnx
+/// ```
+///
+/// To record a .wav file and run this example:
+///
+/// 1. Use a tool such as QuickTime (on macOS) to record audio from your
+///    microphone and save the result.
+///
+/// 2. Use ffmpeg to convert the audio file to a 16 kHz .wav file:
+///
+///    ```
+///    ffmpeg -i saved-file.m4a -ar 16000 output.wav
+///    ```
+///
+/// 3. Run this program on the generated .wav file:
+///
+///    ```
+///    cargo run --release --bin silero_vad silero.rten output.wav
+///    ```
+///
+/// [^1]: <https://github.com/snakers4/silero-vad>
+fn main() -> Result<(), Box<dyn Error>> {
+    let args = parse_args()?;
+
+    let model = Model::load_file(args.model)?;
+    let samples = read_wav_file(&args.wav_file)?;
+
+    let sr_id = model.node_id("sr")?;
+    let input_id = model.node_id("input")?;
+    let state_id = model.node_id("state")?;
+    let state_n_id = model.node_id("stateN")?;
+    let output_id = model.node_id("output")?;
+
+    // Create initial internal state for the model.
+    let mut state: NdTensor<f32, 3> = NdTensor::zeros([2, 1, 128]); // [2, batch, 128]
+    let sample_rate = 16_000;
+    let chunk_duration = 30;
+
+    // Initialize state machine that tracks whether speech is present.
+    let mut speech_state = VadState::new(VadConfig::default());
+    let mut speech_segments = Vec::new();
+    let mut current_segment: Option<Range<Millis>> = None;
+
+    let samples_per_chunk = (sample_rate * chunk_duration) / 1_000;
+    for (i, chunk) in samples.chunks(samples_per_chunk).enumerate() {
+        let padded_chunk: Vec<_> = chunk
+            .iter()
+            .copied()
+            .chain(std::iter::repeat(0.))
+            .take(samples_per_chunk)
+            .collect();
+
+        let [output, next_state] = model.run_n(
+            [
+                (
+                    input_id,
+                    NdTensor::from_data([1, samples_per_chunk], padded_chunk).into(),
+                ),
+                (sr_id, NdTensor::from(sample_rate as i32).into()),
+                (state_id, state.view().into()),
+            ]
+            .into(),
+            [output_id, state_n_id],
+            None,
+        )?;
+
+        let curr_chunk_duration = (chunk_duration * samples_per_chunk) / chunk.len();
+        let start_time = i * chunk_duration;
+        let end_time = start_time + curr_chunk_duration;
+
+        let output: NdTensor<f32, 2> = output.try_into()?;
+        let prob = output
+            .get([0, 0])
+            .copied()
+            .ok_or::<&str>("empty probability output")?;
+        speech_state.update(end_time, prob);
+
+        if let Some(speech_start) = speech_state.speech_start() {
+            if let Some(current_segment) = current_segment.as_mut() {
+                current_segment.end = end_time;
+            } else {
+                current_segment = Some(speech_start..end_time);
+            }
+        } else if let Some(segment) = current_segment {
+            speech_segments.push(segment);
+            current_segment = None;
+        }
+
+        println!(
+            "time: {}..{}ms speech prob: {:.3} chunk speech: {} speaking: {}",
+            start_time,
+            end_time,
+            prob,
+            speech_state.prev_chunk_speaking,
+            speech_state.speech_start().is_some(),
+        );
+
+        state = next_state.try_into()?;
+    }
+    if let Some(current_segment) = current_segment {
+        speech_segments.push(current_segment);
+    }
+
+    for segment in speech_segments {
+        let start_sec = segment.start as f32 / 1000.0;
+        let end_sec = segment.end as f32 / 1000.0;
+        println!("Speech segment: {:.2}..{:.2}", start_sec, end_sec);
+    }
+
+    Ok(())
+}