feat: add text to speech (#69)

Reviewed-on: #69
2026-10-30 07:16:46 +00:00
parent af2a276b1b
commit 23d7017f25
11 changed files with 420 additions and 67 deletions
@@ -2,6 +2,7 @@
 pub mod model;
 #[cfg(not(target_arch = "wasm32"))]
 mod parakeet;
+pub mod speaker;
 #[cfg(target_arch = "wasm32")]
 mod web;

@@ -0,0 +1,88 @@
+use std::time::Duration;
+
+use tts::Tts;
+
+/// Words per minute of an engine at its normal rate, for estimating speech duration.
+const NORMAL_WPM: f32 = 160.0;
+
+/// Extra time past the estimate before speech is considered finished, for backends that
+/// cannot report utterance end.
+pub const END_GRACE: Duration = Duration::from_secs(5);
+
+/// Reads text aloud through the platform's speech backend, created on first use.
+pub struct Speaker {
+    tts: Option<Tts>,
+    end_tx: flume::Sender<()>,
+    end_rx: flume::Receiver<()>,
+}
+
+impl Speaker {
+    pub fn new() -> Self {
+        let (end_tx, end_rx) = flume::unbounded();
+
+        Self {
+            tts: None,
+            end_tx,
+            end_rx,
+        }
+    }
+
+    /// Cloneable stream of utterance-finish events.
+    pub fn end_events(&self) -> flume::Receiver<()> {
+        self.end_rx.clone()
+    }
+
+    /// Speak the text, interrupting what this speaker is already saying.
+    pub fn speak(&mut self, text: &str) -> Result<Duration, tts::Error> {
+        let end_tx = self.end_tx.clone();
+        let tts = match self.tts.as_ref() {
+            Some(tts) => tts,
+            None => self.tts.insert({
+                let tts = Tts::default()?;
+                if let Err(error) = tts.on_utterance_end(move |_| {
+                    // A send error only means the panel dropped its event loop.
+                    let _ = end_tx.send(());
+                }) {
+                    log::warn!("utterance end events unavailable ({error})");
+                }
+                tts
+            }),
+        };
+        tts.speak(text, true)?;
+        Ok(Self::estimate(tts, text))
+    }
+
+    /// Stop speaking, a speaker that never spoke is already silent.
+    pub fn stop(&mut self) -> Result<(), tts::Error> {
+        match self.tts.as_ref() {
+            Some(tts) => tts.stop()?,
+            None => return Ok(()),
+        }
+
+        Ok(())
+    }
+
+    /// Best-effort speaking state, a speaker that never spoke is silent.
+    pub fn speaking(&self) -> bool {
+        self.tts
+            .as_ref()
+            .and_then(|tts| tts.is_speaking().ok())
+            .unwrap_or(false)
+    }
+
+    /// How long speaking the text is expected to take at the engine's current speed.
+    fn estimate(tts: &Tts, text: &str) -> Duration {
+        let words = text.split_whitespace().count().max(1) as f32;
+        let speed = match (tts.get_rate(), tts.normal_rate()) {
+            (Ok(rate), Ok(normal)) if normal > 0. => (rate / normal).clamp(0.2, 4.),
+            _ => 1.,
+        };
+        Duration::from_secs_f32(words / NORMAL_WPM * 60. / speed)
+    }
+}
+
+impl Default for Speaker {
+    fn default() -> Self {
+        Self::new()
+    }
+}