use std::sync::Arc;
use bunsen::{
errors::{
BunsenResult,
ResultContext,
},
kits::speech::asr::{
driver::TranscriptSegment,
voxtral::{
VoxtralConstruct,
driver::{
VoxtralEngine,
VoxtralStreamDriver,
},
},
},
};
use crate::asr::{
Emission,
Transcript,
Word,
millis,
segment_speaker,
speaker_turns,
stream,
};
#[derive(clap::Args, Debug, Default)]
pub struct VoxtralFlags {}
pub fn driver(
bundle: Arc<VoxtralConstruct>,
emission: Emission,
) -> BunsenResult<VoxtralStreamDriver> {
log::info!(
"voxtral: delay {} ms",
bundle.layout.delay_positions() * bundle.layout.config().position_ms
);
let engine = VoxtralEngine::new(bundle)
.context("the model policy's window_seconds")
.as_policy()?;
let driver = VoxtralStreamDriver::new(engine, emission.into())?;
Ok(driver)
}
pub fn transcribe(
driver: &VoxtralStreamDriver,
samples: &[f32],
chunk: usize,
) -> BunsenResult<(Transcript, Vec<TranscriptSegment>)> {
let (segments, _) = stream::stream(driver, samples, chunk)?;
let words = segments
.iter()
.filter_map(|s| {
let text = s.text.as_deref()?.trim();
(!text.is_empty()).then(|| Word {
word: text.to_string(),
start: millis(s.start),
end: millis(s.end),
tokens: s.tokens.clone(),
speaker: segment_speaker(s),
})
})
.collect();
let transcript = Transcript {
text: stream::join_text(&segments),
turns: driver.speakers().map(|_| speaker_turns(&segments)),
language: None,
words: Some(words),
segments: None,
};
Ok((transcript, segments))
}