mirrors/gst-plugins-rs
mirrors/gst-plugins-rs
1
1
Fork 0
mirror of https://gitlab.freedesktop.org/gstreamer/gst-plugins-rs.git synced 2024-11-25 04:51:26 +00:00
Code Issues Projects Releases Wiki Activity

audiornnoise: Attach audio level meta to output buffers

Browse source 
This is useful downstream for processing of audio voice payloads, for
instance feeding a speech recognition library such as Whisper.

Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-rs/-/merge_requests/1231>
This commit is contained in:
Philippe Normand 2023-06-03 10:44:10 +01:00
parent 90e06dc37b
commit 4d9263f932
2 changed files with 36 additions and 12 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

6
audio/audiofx/Cargo.toml
View file

@ -9,9 +9,9 @@ edition = "2021"
rust-version = "1.70" rust-version = "1.70"
[dependencies] [dependencies]
gst = { package = "gstreamer", git = "https://gitlab.freedesktop.org/gstreamer/gstreamer-rs", features = ["v1_16"] } gst = { package = "gstreamer", git = "https://gitlab.freedesktop.org/gstreamer/gstreamer-rs", features = ["v1_20"] }
gst-base = { package = "gstreamer-base", git = "https://gitlab.freedesktop.org/gstreamer/gstreamer-rs", features = ["v1_16"] } gst-base = { package = "gstreamer-base", git = "https://gitlab.freedesktop.org/gstreamer/gstreamer-rs", features = ["v1_20"] }
gst-audio = { package = "gstreamer-audio", git = "https://gitlab.freedesktop.org/gstreamer/gstreamer-rs", features = ["v1_16"] } gst-audio = { package = "gstreamer-audio", git = "https://gitlab.freedesktop.org/gstreamer/gstreamer-rs", features = ["v1_20"] }
anyhow = "1" anyhow = "1"
byte-slice-cast = "1.0" byte-slice-cast = "1.0"
num-traits = "0.2" num-traits = "0.2"

34
audio/audiofx/src/audiornnoise/imp.rs
View file

@ -129,10 +129,13 @@ impl AudioRNNoise {
buffer.set_duration(duration); buffer.set_duration(duration);
buffer.set_pts(pts); buffer.set_pts(pts);
let (level, has_voice) = {
let mut out_map = buffer.map_writable().map_err(|_| gst::FlowError::Error)?; let mut out_map = buffer.map_writable().map_err(|_| gst::FlowError::Error)?;
let out_data = out_map.as_mut_slice_of::<f32>().unwrap(); let out_data = out_map.as_mut_slice_of::<f32>().unwrap();
self.process(state, &settings, in_data, out_data)
};
self.process(state, &settings, in_data, out_data); gst_audio::AudioLevelMeta::add(buffer, level, has_voice);
} }
self.obj().src_pad().push(buffer) self.obj().src_pad().push(buffer)
@ -160,10 +163,13 @@ impl AudioRNNoise {
buffer.set_duration(duration); buffer.set_duration(duration);
buffer.set_pts(pts); buffer.set_pts(pts);
let (level, has_voice) = {
let mut out_map = buffer.map_writable().map_err(|_| gst::FlowError::Error)?; let mut out_map = buffer.map_writable().map_err(|_| gst::FlowError::Error)?;
let out_data = out_map.as_mut_slice_of::<f32>().unwrap(); let out_data = out_map.as_mut_slice_of::<f32>().unwrap();
self.process(state, &settings, in_data, out_data)
};
self.process(state, &settings, in_data, out_data); gst_audio::AudioLevelMeta::add(buffer, level, has_voice);
} }
Ok(GenerateOutputSuccess::Buffer(buffer)) Ok(GenerateOutputSuccess::Buffer(buffer))
@ -175,9 +181,10 @@ impl AudioRNNoise {
settings: &Settings, settings: &Settings,
input_plane: &[f32], input_plane: &[f32],
output_plane: &mut [f32], output_plane: &mut [f32],
) { ) -> (u8, bool) {
let channels = state.in_info.channels() as usize; let channels = state.in_info.channels() as usize;
let size = FRAME_SIZE * channels; let size = FRAME_SIZE * channels;
let mut has_voice = false;
for (out_frame, in_frame) in output_plane.chunks_mut(size).zip(input_plane.chunks(size)) { for (out_frame, in_frame) in output_plane.chunks_mut(size).zip(input_plane.chunks(size)) {
for (index, item) in in_frame.iter().enumerate() { for (index, item) in in_frame.iter().enumerate() {
@ -207,11 +214,15 @@ impl AudioRNNoise {
); );
} }
gst::debug!(CAT, imp: self, "Voice activity: {}", vad); gst::trace!(CAT, imp: self, "Voice activity: {}", vad);
if vad < settings.vad_threshold { if vad < settings.vad_threshold {
out_frame.fill(0.0); out_frame.fill(0.0);
} else { } else {
// Upon voice activity nnoiseless never really reports a 1.0
// VAD, so we use a hardcoded value close to 1.0 here.
if vad >= 0.98 {
has_voice = true;
}
for (index, item) in out_frame.iter_mut().enumerate() { for (index, item) in out_frame.iter_mut().enumerate() {
let channel_index = index % channels; let channel_index = index % channels;
let channel_denoiser = &state.denoisers[channel_index]; let channel_denoiser = &state.denoisers[channel_index];
@ -220,6 +231,19 @@ impl AudioRNNoise {
} }
} }
} }
let rms = output_plane.iter().copied().map(|x| x * x).sum::<f32>();
let level = (20.0 * f32::log10(rms + f32::EPSILON)) as u8;
gst::trace!(
CAT,
imp: self,
"rms: {}, level: {}, has_voice : {} ", rms,
level,
has_voice
);
(level, has_voice)
} }
} }