Skip to content

armgit / ente

Public Imported from GitHub
0

💚 End-to-end encrypted cloud for everything.

[ensu] Match the pinned transcript spacing

laurenspriem9fd3838bb35e
diff --git a/rust/crates/ensu/src/transcription/parakeet.rs b/rust/crates/ensu/src/transcription/parakeet.rs
index af674bf203..1fb161212d 100644
--- a/rust/crates/ensu/src/transcription/parakeet.rs
+++ b/rust/crates/ensu/src/transcription/parakeet.rs
@@ -5,6 +5,7 @@ use ort::{
session::{Session, SessionOutputs, builder::GraphOptimizationLevel},
value::{DynValue, TensorRef},
};
+use regex::{Captures, regex};
use super::{Result, TranscriptionError};
@@ -230,11 +231,11 @@ fn render(emissions: &[Emission], vocabulary: &[String]) -> Transcript {
word.end = end;
}
}
- transcript.text = transcript
- .text
- .split_whitespace()
- .collect::<Vec<_>>()
- .join(" ");
+ transcript.text = regex!(r"\A\s|\s\B|(\s)\b")
+ .replace_all(&transcript.text, |captures: &Captures<'_>| {
+ if captures.get(1).is_some() { " " } else { "" }
+ })
+ .into_owned();
for word in &mut transcript.words {
word.start = (word.start - 0.25).max(0.0);
word.end = (word.end - 0.25).max(0.0);
@@ -329,4 +330,31 @@ mod tests {
assert_eq!(continuation.words[0].text, "Hello世界!");
assert_eq!(continuation.words[0].end, 7.0 * FRAME_SECONDS + 0.05 - 0.25);
}
+
+ #[test]
+ fn text_drops_spaces_before_non_word_tokens_without_changing_words() {
+ let vocabulary = vec![
+ "▁cost".into(),
+ "▁$".into(),
+ "5".into(),
+ "▁".into(),
+ ",".into(),
+ "▁¿".into(),
+ "qué".into(),
+ ];
+ let emissions: Vec<_> = (0..vocabulary.len())
+ .map(|token| Emission {
+ token,
+ frame: token,
+ })
+ .collect();
+ let transcript = render(&emissions, &vocabulary);
+ assert_eq!(transcript.text, "cost$5,¿qué");
+ let words: Vec<_> = transcript
+ .words
+ .iter()
+ .map(|word| word.text.as_str())
+ .collect();
+ assert_eq!(words, ["cost", "$5,", "¿qué"]);
+ }
}