net/aws/transcriber: use two queues for sending transcript items

* A queue dedicated to transcript items not intended for translation.
* A queue dedicated to transcript items intended for translation. The items are
  enqueued after a separator is detected or translate-lookahead was reached.

Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-rs/-/merge_requests/1137>

3 files changed, 374 insertions, 324 deletions

diff --git a/net/aws/src/transcriber/imp.rs b/net/aws/src/transcriber/imp.rs
index 0c99c760..cc3f8ad2 100644
--- a/net/aws/src/transcriber/imp.rs
+++ b/net/aws/src/transcriber/imp.rs
@@ -29,12 +29,12 @@ use futures::prelude::*;
 use tokio::{runtime, sync::broadcast, task};
 
 use std::collections::{BTreeSet, VecDeque};
-use std::sync::Mutex;
+use std::sync::{Arc, Mutex};
 
 use once_cell::sync::Lazy;
 
-use super::transcribe::{TranscriberLoop, TranscriptEvent, TranscriptItem, TranscriptionSettings};
-use super::translate::{TranslateLoop, TranslateQueue, TranslatedItem};
+use super::transcribe::{TranscriberSettings, TranscriberStream, TranscriptEvent, TranscriptItem};
+use super::translate::{TranslateLoop, TranslatedItem};
 use super::{
     AwsTranscriberResultStability, AwsTranscriberVocabularyFilterMethod,
     TranslationTokenizationMethod, CAT,
@@ -148,6 +148,7 @@ struct State {
     srcpads: BTreeSet<super::TranslateSrcPad>,
     pad_serial: u32,
     seqnum: gst::Seqnum,
+    start_time: Option<gst::ClockTime>,
 }
 
 impl Default for State {
@@ -158,6 +159,7 @@ impl Default for State {
             srcpads: Default::default(),
             pad_serial: 0,
             seqnum: gst::Seqnum::next(),
+            start_time: None,
         }
     }
 }
@@ -168,7 +170,9 @@ pub struct Transcriber {
     settings: Mutex<Settings>,
     state: Mutex<State>,
     pub(super) aws_config: Mutex<Option<aws_config::SdkConfig>>,
-    // sender to broadcast transcript items to the translate src pads.
+    // sender to broadcast transcript items to the src pads for translation.
+    transcript_event_for_translate_tx: broadcast::Sender<TranscriptEvent>,
+    // sender to broadcast transcript items to the src pads, not intended for translation.
     transcript_event_tx: broadcast::Sender<TranscriptEvent>,
 }
 
@@ -276,7 +280,10 @@ impl Transcriber {
     ) -> Result<gst::FlowSuccess, gst::FlowError> {
         gst::log!(CAT, obj: pad, "Handling {buffer:?}");
 
-        self.ensure_connection();
+        self.ensure_connection().map_err(|err| {
+            gst::element_imp_error!(self, gst::StreamError::Failed, ["Streaming failed: {err}"]);
+            gst::FlowError::Error
+        })?;
 
         let Some(mut buffer_tx) = self.state.lock().unwrap().buffer_tx.take() else {
             gst::log!(CAT, obj: pad, "Flushing");
@@ -292,12 +299,82 @@ impl Transcriber {
 
         Ok(gst::FlowSuccess::Ok)
     }
+}
+
+#[derive(Default)]
+struct TranslateQueue {
+    items: VecDeque<TranscriptItem>,
+}
+
+impl TranslateQueue {
+    fn is_empty(&self) -> bool {
+        self.items.is_empty()
+    }
+
+    /// Pushes the provided item.
+    ///
+    /// Returns `Some(..)` if items are ready for translation.
+    fn push(&mut self, transcript_item: &TranscriptItem) -> Option<Vec<TranscriptItem>> {
+        // Keep track of the item individually so we can schedule translation precisely.
+        self.items.push_back(transcript_item.clone());
+
+        if transcript_item.is_punctuation {
+            // This makes it a good chunk for translation.
+            // Concatenate as a single item for translation
+
+            return Some(self.items.drain(..).collect());
+        }
+
+        // Regular case: no separator detected, don't push transcript items
+        // to translation now. They will be pushed either if a punctuation
+        // is found or of a `dequeue()` is requested.
+
+        None
+    }
+
+    /// Dequeues items from the specified `deadline` up to `lookahead`.
+    ///
+    /// Returns `Some(..)` if some items match the criteria.
+    fn dequeue(
+        &mut self,
+        latency: gst::ClockTime,
+        threshold: gst::ClockTime,
+        lookahead: gst::ClockTime,
+    ) -> Option<Vec<TranscriptItem>> {
+        let first_pts = self.items.front()?.pts;
+        if first_pts + latency > threshold {
+            // First item is too early to be sent to translation now
+            // we can wait for more items to accumulate.
+            return None;
+        }
+
+        // Can't wait any longer to send the first item to translation
+        // Try to get up to lookahead worth of items to improve translation accuracy
+        let limit = first_pts + lookahead;
+
+        let mut items_acc = vec![self.items.pop_front().unwrap()];
+        while let Some(item) = self.items.front() {
+            if item.pts > limit {
+                break;
+            }
+
+            items_acc.push(self.items.pop_front().unwrap());
+        }
 
-    fn ensure_connection(&self) {
+        Some(items_acc)
+    }
+
+    fn drain(&mut self) -> impl Iterator<Item = TranscriptItem> + '_ {
+        self.items.drain(..)
+    }
+}
+
+impl Transcriber {
+    fn ensure_connection(&self) -> Result<(), gst::ErrorMessage> {
         let mut state = self.state.lock().unwrap();
 
         if state.buffer_tx.is_some() {
-            return;
+            return Ok(());
         }
 
         let settings = self.settings.lock().unwrap();
@@ -306,21 +383,116 @@ impl Transcriber {
         let s = in_caps.structure(0).unwrap();
         let sample_rate = s.get::<i32>("rate").unwrap();
 
-        let transcription_settings = TranscriptionSettings::from(&settings, sample_rate);
+        let transcription_settings = TranscriberSettings::from(&settings, sample_rate);
 
         let (buffer_tx, buffer_rx) = mpsc::channel(1);
 
-        let transcriber_loop = TranscriberLoop::new(
+        let _enter = RUNTIME.enter();
+        let mut transcriber_stream = futures::executor::block_on(TranscriberStream::try_new(
             self,
             transcription_settings,
             settings.lateness,
             buffer_rx,
-            self.transcript_event_tx.clone(),
-        );
-        let transcriber_loop_handle = RUNTIME.spawn(transcriber_loop.run());
+        ))?;
+
+        // Latency budget for an item to be pushed to stream on time
+        // Margin:
+        // - 2 * GRANULARITY: to make sure we don't push items up to GRANULARITY late.
+        // - 1 * GRANULARITY: extra margin to account for additional overheads.
+        let latency = settings.transcribe_latency.saturating_sub(3 * GRANULARITY);
+        let translate_lookahead = settings.translate_lookahead;
+        let mut translate_queue = TranslateQueue::default();
+        let imp = self.ref_counted();
+        let transcriber_loop_handle = RUNTIME.spawn(async move {
+            loop {
+                // This is to make sure we send items on a timely basis or at least Gap events.
+                let timeout = tokio::time::sleep(GRANULARITY.into()).fuse();
+                futures::pin_mut!(timeout);
+
+                let transcriber_next = transcriber_stream.next().fuse();
+                futures::pin_mut!(transcriber_next);
+
+                // `transcriber_next` takes precedence over `timeout`
+                // because we don't want to loose any incoming items.
+                let res = futures::select_biased! {
+                    event = transcriber_next => Some(event?),
+                    _ = timeout => None,
+                };
+
+                use TranscriptEvent::*;
+                match res {
+                    None => (),
+                    Some(Items(items)) => {
+                        if imp.transcript_event_tx.receiver_count() > 0 {
+                            let _ = imp.transcript_event_tx.send(Items(items.clone()));
+                        }
+
+                        if imp.transcript_event_for_translate_tx.receiver_count() > 0 {
+                            for item in items.iter() {
+                                if let Some(items_to_translate) = translate_queue.push(item) {
+                                    let _ = imp
+                                        .transcript_event_for_translate_tx
+                                        .send(Items(items_to_translate.into()));
+                                }
+                            }
+                        }
+                    }
+                    Some(Eos) => {
+                        gst::debug!(CAT, imp: imp, "Transcriber loop sending EOS");
+
+                        if imp.transcript_event_tx.receiver_count() > 0 {
+                            let _ = imp.transcript_event_tx.send(Eos);
+                        }
+
+                        if imp.transcript_event_for_translate_tx.receiver_count() > 0 {
+                            let items_to_translate: Vec<TranscriptItem> =
+                                translate_queue.drain().collect();
+                            let _ = imp
+                                .transcript_event_for_translate_tx
+                                .send(Items(items_to_translate.into()));
+
+                            let _ = imp.transcript_event_for_translate_tx.send(Eos);
+                        }
+
+                        break;
+                    }
+                }
+
+                if imp.transcript_event_for_translate_tx.receiver_count() > 0 {
+                    // Check if we need to push items for translation
+
+                    let Some((start_time, now)) = imp.get_start_time_and_now() else {
+                        continue;
+                    };
+
+                    if !translate_queue.is_empty() {
+                        let threshold = now - start_time;
+
+                        if let Some(items_to_translate) =
+                            translate_queue.dequeue(latency, threshold, translate_lookahead)
+                        {
+                            gst::debug!(
+                                CAT,
+                                imp: imp,
+                                "Forcing to translation (threshold {threshold}): {items_to_translate:?}"
+                            );
+                            let _ = imp
+                                .transcript_event_for_translate_tx
+                                .send(Items(items_to_translate.into()));
+                        }
+                    }
+                }
+            }
+
+            gst::debug!(CAT, imp: imp, "Exiting transcriber loop");
+
+            Ok(())
+        });
 
         state.transcriber_loop_handle = Some(transcriber_loop_handle);
         state.buffer_tx = Some(buffer_tx);
+
+        Ok(())
     }
 
     fn prepare(&self) -> Result<(), gst::ErrorMessage> {
@@ -382,6 +554,18 @@ impl Transcriber {
         }
         gst::info!(CAT, imp: self, "Unprepared");
     }
+
+    fn get_start_time_and_now(&self) -> Option<(gst::ClockTime, gst::ClockTime)> {
+        let now = self.obj().current_running_time()?;
+
+        let mut state = self.state.lock().unwrap();
+
+        if state.start_time.is_none() {
+            state.start_time = Some(now);
+        }
+
+        Some((state.start_time.unwrap(), now))
+    }
 }
 
 #[glib::object_subclass]
@@ -438,6 +622,7 @@ impl ObjectSubclass for Transcriber {
         // Setting the channel capacity so that a TranslateSrcPad that would lag
         // behind for some reasons get a chance to catch-up without loosing items.
         // Receiver will be created by subscribing to sender later.
+        let (transcript_event_for_translate_tx, _) = broadcast::channel(128);
         let (transcript_event_tx, _) = broadcast::channel(128);
 
         Self {
@@ -446,6 +631,7 @@ impl ObjectSubclass for Transcriber {
             settings: Default::default(),
             state: Default::default(),
             aws_config: Default::default(),
+            transcript_event_for_translate_tx,
             transcript_event_tx,
         }
     }
@@ -876,51 +1062,93 @@ struct TranslationPadTask {
     elem: super::Transcriber,
     transcript_event_rx: broadcast::Receiver<TranscriptEvent>,
     needs_translate: bool,
-    translate_queue: TranslateQueue,
     translate_loop_handle: Option<task::JoinHandle<Result<(), gst::ErrorMessage>>>,
-    to_translate_tx: Option<mpsc::Sender<Vec<TranscriptItem>>>,
+    to_translate_tx: Option<mpsc::Sender<Arc<Vec<TranscriptItem>>>>,
     from_translate_rx: Option<mpsc::Receiver<Vec<TranslatedItem>>>,
-    translate_latency: gst::ClockTime,
-    translate_lookahead: gst::ClockTime,
     send_events: bool,
     output_items: VecDeque<OutputItem>,
     our_latency: gst::ClockTime,
     seqnum: gst::Seqnum,
     send_eos: bool,
     pending_translations: usize,
-    start_time: Option<gst::ClockTime>,
 }
 
 impl TranslationPadTask {
-    fn try_new(
+    async fn try_new(
         pad: &TranslateSrcPad,
         elem: super::Transcriber,
-        transcript_event_rx: broadcast::Receiver<TranscriptEvent>,
     ) -> Result<TranslationPadTask, gst::ErrorMessage> {
-        let mut this = TranslationPadTask {
+        let mut translation_loop = None;
+        let mut translate_loop_handle = None;
+        let mut to_translate_tx = None;
+        let mut from_translate_rx = None;
+
+        let (our_latency, transcript_event_rx, needs_translate);
+
+        {
+            let elem_imp = elem.imp();
+            let elem_settings = elem_imp.settings.lock().unwrap();
+
+            let pad_settings = pad.settings.lock().unwrap();
+
+            our_latency = TranslateSrcPad::our_latency(&elem_settings, &pad_settings);
+            if our_latency + elem_settings.lateness <= 2 * GRANULARITY {
+                let err = format!(
+                    "total latency + lateness must be greater than {}",
+                    2 * GRANULARITY
+                );
+                gst::error!(CAT, imp: pad, "{err}");
+                return Err(gst::error_msg!(gst::LibraryError::Settings, ["{err}"]));
+            }
+
+            needs_translate = TranslateSrcPad::needs_translation(
+                &elem_settings.language_code,
+                pad_settings.language_code.as_deref(),
+            );
+
+            if needs_translate {
+                let (to_loop_tx, to_loop_rx) = mpsc::channel(64);
+                let (from_loop_tx, from_loop_rx) = mpsc::channel(64);
+
+                translation_loop = Some(TranslateLoop::new(
+                    elem_imp,
+                    pad,
+                    &elem_settings.language_code,
+                    pad_settings.language_code.as_deref().unwrap(),
+                    pad_settings.tokenization_method,
+                    to_loop_rx,
+                    from_loop_tx,
+                ));
+
+                to_translate_tx = Some(to_loop_tx);
+                from_translate_rx = Some(from_loop_rx);
+
+                transcript_event_rx = elem_imp.transcript_event_for_translate_tx.subscribe();
+            } else {
+                transcript_event_rx = elem_imp.transcript_event_tx.subscribe();
+            }
+        }
+
+        if let Some(translation_loop) = translation_loop {
+            translation_loop.check_language().await?;
+            translate_loop_handle = Some(RUNTIME.spawn(translation_loop.run()));
+        }
+
+        Ok(TranslationPadTask {
             pad: pad.ref_counted(),
             elem,
             transcript_event_rx,
-            needs_translate: false,
-            translate_queue: TranslateQueue::default(),
-            translate_loop_handle: None,
-            to_translate_tx: None,
-            from_translate_rx: None,
-            translate_latency: DEFAULT_TRANSLATE_LATENCY,
-            translate_lookahead: DEFAULT_TRANSLATE_LOOKAHEAD,
+            needs_translate,
+            translate_loop_handle,
+            to_translate_tx,
+            from_translate_rx,
             send_events: true,
             output_items: VecDeque::new(),
-            our_latency: DEFAULT_TRANSCRIBE_LATENCY,
+            our_latency,
             seqnum: gst::Seqnum::next(),
             send_eos: false,
             pending_translations: 0,
-            start_time: None,
-        };
-
-        let _enter_guard = RUNTIME.enter();
-        futures::executor::block_on(this.init_translate())?;
-
-        Ok(this)
+        })
     }
 }
 
@@ -958,11 +1186,9 @@ impl TranslationPadTask {
         let transcript_event_rx = self.transcript_event_rx.recv().fuse();
         futures::pin_mut!(transcript_event_rx);
 
-        // `timeout` takes precedence over `transcript_events` reception
-        // because we may need to `dequeue` `items` or push a `Gap` event
-        // before current latency budget is exhausted.
+        // `transcript_event_rx` takes precedence over `timeout`
+        // because we don't want to loose any incoming items.
         futures::select_biased! {
-            _ = timeout => (),
             items_res = transcript_event_rx => {
                 use TranscriptEvent::*;
                 use broadcast::error::RecvError;
@@ -983,6 +1209,7 @@ impl TranslationPadTask {
                     }
                 }
             }
+            _ = timeout => (),
         }
 
         Ok(())
@@ -999,121 +1226,100 @@ impl TranslationPadTask {
             return Err(gst::error_msg!(gst::StreamError::Failed, ["{ERR}"]));
         }
 
-        let transcript_items = {
+        let items_to_translate = {
             // This is to make sure we send items on a timely basis or at least Gap events.
             let timeout = tokio::time::sleep(GRANULARITY.into()).fuse();
             futures::pin_mut!(timeout);
 
-            let from_translate_rx = self
-                .from_translate_rx
-                .as_mut()
-                .expect("from_translation chan must be available in translation mode");
-
             let transcript_event_rx = self.transcript_event_rx.recv().fuse();
             futures::pin_mut!(transcript_event_rx);
 
-            // `timeout` takes precedence over `transcript_events` reception
-            // because we may need to `dequeue` `items` or push a `Gap` event
-            // before current latency budget is exhausted.
+            // `transcript_event_rx` takes precedence over `timeout`
+            // because we don't want to loose any incoming items.
             futures::select_biased! {
-                _ = timeout => return Ok(()),
-                translated_items = from_translate_rx.next() => {
-                    let Some(translated_items) = translated_items else {
-                        const ERR: &str = "translation chan terminated";
-                        gst::debug!(CAT, imp: self.pad, "{ERR}");
-                        return Err(gst::error_msg!(gst::StreamError::Failed, ["{ERR}"]));
-                    };
-
-                    self.output_items.extend(translated_items.into_iter().map(Into::into));
-                    self.pending_translations = self.pending_translations.saturating_sub(1);
-
-                    return Ok(());
-                }
                 items_res = transcript_event_rx => {
                     use TranscriptEvent::*;
                     use broadcast::error::RecvError;
                     match items_res {
-                        Ok(Items(transcript_items)) => transcript_items,
+                        Ok(Items(items_to_translate)) => Some(items_to_translate),
                         Ok(Eos) => {
                             gst::debug!(CAT, imp: self.pad, "Got eos");
                             self.send_eos = true;
-                            return Ok(());
+                            None
                         }
                         Err(RecvError::Lagged(nb_msg)) => {
                             gst::warning!(CAT, imp: self.pad, "Missed {nb_msg} transcript sets");
-                            return Ok(());
+                            None
                         }
                         Err(RecvError::Closed) => {
                             gst::debug!(CAT, imp: self.pad, "Transcript chan terminated: setting eos");
                             self.send_eos = true;
-                            return Ok(());
+                            None
                         }
                     }
                 }
+                _ = timeout => None,
             }
         };
 
-        for items in transcript_items.iter() {
-            if let Some(items_to_translate) = self.translate_queue.push(items) {
-                self.send_for_translation(items_to_translate).await?;
+        if let Some(items_to_translate) = items_to_translate {
+            if !items_to_translate.is_empty() {
+                let res = self
+                    .to_translate_tx
+                    .as_mut()
+                    .expect("to_translation chan must be available in translation mode")
+                    .send(items_to_translate)
+                    .await;
+
+                if res.is_err() {
+                    const ERR: &str = "to_translation chan terminated";
+                    gst::debug!(CAT, imp: self.pad, "{ERR}");
+                    return Err(gst::error_msg!(gst::StreamError::Failed, ["{ERR}"]));
+                }
+
+                self.pending_translations += 1;
             }
         }
 
-        Ok(())
-    }
+        // Check pending translated items
+        let from_translate_rx = self
+            .from_translate_rx
+            .as_mut()
+            .expect("from_translation chan must be available in translation mode");
 
-    async fn dequeue_for_translation(
-        &mut self,
-        start_time: gst::ClockTime,
-        now: gst::ClockTime,
-    ) -> Result<(), gst::ErrorMessage> {
-        if !self.translate_queue.is_empty() {
-            // Latency budget for an item to be pushed to stream on time
-            // Margin:
-            // - 2 * GRANULARITY: to make sure we don't push items up to GRANULARITY late.
-            // - 1 * GRANULARITY: extra margin to account for additional overheads.
-            let latency = self.our_latency.saturating_sub(3 * GRANULARITY);
-
-            // Estimated time of arrival for an item sent to translation now.
-            // (in transcript item ts base)
-            let translation_eta = now + self.translate_latency - start_time;
-
-            if let Some(items_to_translate) =
-                self.translate_queue
-                    .dequeue(latency, translation_eta, self.translate_lookahead)
-            {
-                gst::debug!(CAT, imp: self.pad, "Forcing to translation: {items_to_translate:?}");
-                self.send_for_translation(items_to_translate).await?;
-            }
+        while let Ok(translated_items) = from_translate_rx.try_next() {
+            let Some(translated_items) = translated_items else {
+                const ERR: &str = "translation chan terminated";
+                gst::debug!(CAT, imp: self.pad, "{ERR}");
+                return Err(gst::error_msg!(gst::StreamError::Failed, ["{ERR}"]));
+            };
+
+            self.output_items
+                .extend(translated_items.into_iter().map(Into::into));
+            self.pending_translations = self.pending_translations.saturating_sub(1);
         }
 
         Ok(())
     }
 
     async fn dequeue(&mut self) -> bool {
-        let (now, start_time, mut last_position, mut discont_pending);
-        {
-            let mut pad_state = self.pad.state.lock().unwrap();
-
-            let Some(cur_rt) = self.elem.current_running_time() else {
-                // Wait for the clock to be available
-                return true;
-            };
-            now = cur_rt;
+        let Some((start_time, now)) = self.elem.imp().get_start_time_and_now() else {
+            // Wait for the clock to be available
+            return true;
+        };
 
-            if self.start_time.is_none() {
-                self.start_time = Some(now);
-                pad_state.out_segment.set_position(now);
-            }
+        let (mut last_position, mut discont_pending) = {
+            let mut state = self.pad.state.lock().unwrap();
 
-            start_time = self.start_time.unwrap();
-            last_position = pad_state.out_segment.position().unwrap();
-            discont_pending = pad_state.discont_pending;
-        }
+            let last_position = if let Some(pos) = state.out_segment.position() {
+                pos
+            } else {
+                state.out_segment.set_position(start_time);
+                start_time
+            };
 
-        if self.needs_translate && self.dequeue_for_translation(start_time, now).await.is_err() {
-            return false;
-        }
+            (last_position, state.discont_pending)
+        };
 
         /* First, check our pending buffers */
         while let Some(item) = self.output_items.front() {
@@ -1206,11 +1412,7 @@ impl TranslationPadTask {
             }
         }
 
-        if self.send_eos
-            && self.pending_translations == 0
-            && self.output_items.is_empty()
-            && self.translate_queue.is_empty()
-        {
+        if self.send_eos && self.pending_translations == 0 && self.output_items.is_empty() {
             /* We're EOS, we can pause and exit early */
             let _ = self.pad.obj().pause_task();
 
@@ -1261,28 +1463,6 @@ impl TranslationPadTask {
         true
     }
 
-    async fn send_for_translation(
-        &mut self,
-        transcript_items: Vec<TranscriptItem>,
-    ) -> Result<(), gst::ErrorMessage> {
-        let res = self
-            .to_translate_tx
-            .as_mut()
-            .expect("to_translation chan must be available in translation mode")
-            .send(transcript_items)
-            .await;
-
-        if res.is_err() {
-            const ERR: &str = "to_translation chan terminated";
-            gst::debug!(CAT, imp: self.pad, "{ERR}");
-            return Err(gst::error_msg!(gst::StreamError::Failed, ["{ERR}"]));
-        }
-
-        self.pending_translations += 1;
-
-        Ok(())
-    }
-
     fn ensure_init_events(&mut self) -> Result<(), gst::ErrorMessage> {
         if !self.send_events {
             return Ok(());
@@ -1332,62 +1512,6 @@ impl TranslationPadTask {
     }
 }
 
-impl TranslationPadTask {
-    async fn init_translate(&mut self) -> Result<(), gst::ErrorMessage> {
-        let mut translation_loop = None;
-
-        {
-            let elem_imp = self.elem.imp();
-            let elem_settings = elem_imp.settings.lock().unwrap();
-
-            let pad_settings = self.pad.settings.lock().unwrap();
-
-            self.our_latency = TranslateSrcPad::our_latency(&elem_settings, &pad_settings);
-            if self.our_latency + elem_settings.lateness <= 2 * GRANULARITY {
-                let err = format!(
-                    "total latency + lateness must be greater than {}",
-                    2 * GRANULARITY
-                );
-                gst::error!(CAT, imp: self.pad, "{err}");
-                return Err(gst::error_msg!(gst::LibraryError::Settings, ["{err}"]));
-            }
-
-            self.translate_latency = elem_settings.translate_latency;
-            self.translate_lookahead = elem_settings.translate_lookahead;
-
-            self.needs_translate = TranslateSrcPad::needs_translation(
-                &elem_settings.language_code,
-                pad_settings.language_code.as_deref(),
-            );
-
-            if self.needs_translate {
-                let (to_translate_tx, to_translate_rx) = mpsc::channel(64);
-                let (from_translate_tx, from_translate_rx) = mpsc::channel(64);
-
-                translation_loop = Some(TranslateLoop::new(
-                    elem_imp,
-                    &self.pad,
-                    &elem_settings.language_code,
-                    pad_settings.language_code.as_deref().unwrap(),
-                    pad_settings.tokenization_method,
-                    to_translate_rx,
-                    from_translate_tx,
-                ));
-
-                self.to_translate_tx = Some(to_translate_tx);
-                self.from_translate_rx = Some(from_translate_rx);
-            }
-        }
-
-        if let Some(translation_loop) = translation_loop {
-            translation_loop.check_language().await?;
-            self.translate_loop_handle = Some(RUNTIME.spawn(translation_loop.run()));
-        }
-
-        Ok(())
-    }
-}
-
 #[derive(Debug)]
 struct TranslationPadState {
     discont_pending: bool,
@@ -1422,8 +1546,8 @@ impl TranslateSrcPad {
         gst::debug!(CAT, imp: self, "Starting task");
 
         let elem = self.parent();
-        let transcript_event_rx = elem.imp().transcript_event_tx.subscribe();
-        let mut pad_task = TranslationPadTask::try_new(self, elem, transcript_event_rx)
+        let _enter = RUNTIME.enter();
+        let mut pad_task = futures::executor::block_on(TranslationPadTask::try_new(self, elem))
             .map_err(|err| gst::loggable_error!(CAT, format!("Failed to start pad task {err}")))?;
 
         let imp = self.ref_counted();
diff --git a/net/aws/src/transcriber/transcribe.rs b/net/aws/src/transcriber/transcribe.rs
index 97301380..d094d87a 100644
--- a/net/aws/src/transcriber/transcribe.rs
+++ b/net/aws/src/transcriber/transcribe.rs
@@ -15,7 +15,6 @@ use aws_sdk_transcribestreaming::model;
 
 use futures::channel::mpsc;
 use futures::prelude::*;
-use tokio::sync::broadcast;
 
 use std::sync::Arc;
 
@@ -23,7 +22,7 @@ use super::imp::{Settings, Transcriber};
 use super::CAT;
 
 #[derive(Debug)]
-pub struct TranscriptionSettings {
+pub struct TranscriberSettings {
     lang_code: model::LanguageCode,
     sample_rate: i32,
     vocabulary: Option<String>,
@@ -33,9 +32,9 @@ pub struct TranscriptionSettings {
     results_stability: model::PartialResultsStability,
 }
 
-impl TranscriptionSettings {
+impl TranscriberSettings {
     pub(super) fn from(settings: &Settings, sample_rate: i32) -> Self {
-        TranscriptionSettings {
+        TranscriberSettings {
             lang_code: settings.language_code.as_str().into(),
             sample_rate,
             vocabulary: settings.vocabulary.clone(),
@@ -83,43 +82,30 @@ impl From<Vec<TranscriptItem>> for TranscriptEvent {
     }
 }
 
-pub struct TranscriberLoop {
+pub struct TranscriberStream {
     imp: glib::subclass::ObjectImplRef<Transcriber>,
-    client: aws_transcribe::Client,
-    settings: Option<TranscriptionSettings>,
+    output: aws_transcribe::output::StartStreamTranscriptionOutput,
     lateness: gst::ClockTime,
-    buffer_rx: Option<mpsc::Receiver<gst::Buffer>>,
-    transcript_items_tx: broadcast::Sender<TranscriptEvent>,
     partial_index: usize,
 }
 
-impl TranscriberLoop {
-    pub fn new(
+impl TranscriberStream {
+    pub async fn try_new(
         imp: &Transcriber,
-        settings: TranscriptionSettings,
+        settings: TranscriberSettings,
         lateness: gst::ClockTime,
         buffer_rx: mpsc::Receiver<gst::Buffer>,
-        transcript_items_tx: broadcast::Sender<TranscriptEvent>,
-    ) -> Self {
-        let aws_config = imp.aws_config.lock().unwrap();
-        let aws_config = aws_config
-            .as_ref()
-            .expect("aws_config must be initialized at this stage");
-
-        TranscriberLoop {
-            imp: imp.ref_counted(),
-            client: aws_transcribe::Client::new(aws_config),
-            settings: Some(settings),
-            lateness,
-            buffer_rx: Some(buffer_rx),
-            transcript_items_tx,
-            partial_index: 0,
-        }
-    }
+    ) -> Result<Self, gst::ErrorMessage> {
+        let client = {
+            let aws_config = imp.aws_config.lock().unwrap();
+            let aws_config = aws_config
+                .as_ref()
+                .expect("aws_config must be initialized at this stage");
+            aws_transcribe::Client::new(aws_config)
+        };
 
-    pub async fn run(mut self) -> Result<(), gst::ErrorMessage> {
         // Stream the incoming buffers chunked
-        let chunk_stream = self.buffer_rx.take().unwrap().flat_map(move |buffer: gst::Buffer| {
+        let chunk_stream = buffer_rx.flat_map(move |buffer: gst::Buffer| {
             async_stream::stream! {
                 let data = buffer.map_readable().unwrap();
                 use aws_transcribe::{model::{AudioEvent, AudioStream}, types::Blob};
@@ -129,9 +115,7 @@ impl TranscriberLoop {
             }
         });
 
-        let settings = self.settings.take().unwrap();
-        let mut transcribe_builder = self
-            .client
+        let mut transcribe_builder = client
             .start_stream_transcription()
             .language_code(settings.lang_code)
             .media_sample_rate_hertz(settings.sample_rate)
@@ -147,26 +131,42 @@ impl TranscriberLoop {
                 .vocabulary_filter_method(settings.vocabulary_filter_method);
         }
 
-        let mut output = transcribe_builder
+        let output = transcribe_builder
             .audio_stream(chunk_stream.into())
             .send()
             .await
             .map_err(|err| {
                 let err = format!("Transcribe ws init error: {err}");
-                gst::error!(CAT, imp: self.imp, "{err}");
+                gst::error!(CAT, imp: imp, "{err}");
                 gst::error_msg!(gst::LibraryError::Init, ["{err}"])
             })?;
 
-        while let Some(event) = output
-            .transcript_result_stream
-            .recv()
-            .await
-            .map_err(|err| {
-                let err = format!("Transcribe ws stream error: {err}");
-                gst::error!(CAT, imp: self.imp, "{err}");
-                gst::error_msg!(gst::LibraryError::Failed, ["{err}"])
-            })?
-        {
+        Ok(TranscriberStream {
+            imp: imp.ref_counted(),
+            output,
+            lateness,
+            partial_index: 0,
+        })
+    }
+
+    pub async fn next(&mut self) -> Result<TranscriptEvent, gst::ErrorMessage> {
+        loop {
+            let event = self
+                .output
+                .transcript_result_stream
+                .recv()
+                .await
+                .map_err(|err| {
+                    let err = format!("Transcribe ws stream error: {err}");
+                    gst::error!(CAT, imp: self.imp, "{err}");
+                    gst::error_msg!(gst::LibraryError::Failed, ["{err}"])
+                })?;
+
+            let Some(event) = event else {
+                gst::debug!(CAT, imp: self.imp, "Transcriber loop sending EOS");
+                return Ok(TranscriptEvent::Eos);
+            };
+
             if let model::TranscriptResultStream::TranscriptEvent(transcript_evt) = event {
                 let mut ready_items = None;
 
@@ -188,10 +188,7 @@ impl TranscriberLoop {
                 }
 
                 if let Some(ready_items) = ready_items {
-                    if self.transcript_items_tx.send(ready_items.into()).is_err() {
-                        gst::debug!(CAT, imp: self.imp, "No transcript items receivers");
-                        break;
-                    }
+                    return Ok(ready_items.into());
                 }
             } else {
                 gst::warning!(
@@ -201,13 +198,6 @@ impl TranscriberLoop {
                 )
             }
         }
-
-        gst::debug!(CAT, imp: self.imp, "Transcriber loop sending EOS");
-        let _ = self.transcript_items_tx.send(TranscriptEvent::Eos);
-
-        gst::debug!(CAT, imp: self.imp, "Exiting transcriber loop");
-
-        Ok(())
     }
 
     /// Builds a list from the provided stable items.
diff --git a/net/aws/src/transcriber/translate.rs b/net/aws/src/transcriber/translate.rs
index b944e001..71f43aef 100644
--- a/net/aws/src/transcriber/translate.rs
+++ b/net/aws/src/transcriber/translate.rs
@@ -14,7 +14,7 @@ use aws_sdk_translate as aws_translate;
 use futures::channel::mpsc;
 use futures::prelude::*;
 
-use std::collections::VecDeque;
+use std::sync::Arc;
 
 use super::imp::TranslateSrcPad;
 use super::transcribe::TranscriptItem;
@@ -40,77 +40,13 @@ impl From<&TranscriptItem> for TranslatedItem {
     }
 }
 
-#[derive(Default)]
-pub struct TranslateQueue {
-    items: VecDeque<TranscriptItem>,
-}
-
-impl TranslateQueue {
-    pub fn is_empty(&self) -> bool {
-        self.items.is_empty()
-    }
-
-    /// Pushes the provided item.
-    ///
-    /// Returns `Some(..)` if items are ready for translation.
-    pub fn push(&mut self, transcript_item: &TranscriptItem) -> Option<Vec<TranscriptItem>> {
-        // Keep track of the item individually so we can schedule translation precisely.
-        self.items.push_back(transcript_item.clone());
-
-        if transcript_item.is_punctuation {
-            // This makes it a good chunk for translation.
-            // Concatenate as a single item for translation
-
-            return Some(self.items.drain(..).collect());
-        }
-
-        // Regular case: no separator detected, don't push transcript items
-        // to translation now. They will be pushed either if a punctuation
-        // is found or of a `dequeue()` is requested.
-
-        None
-    }
-
-    /// Dequeues items from the specified `deadline` up to `lookahead`.
-    ///
-    /// Returns `Some(..)` if some items match the criteria.
-    pub fn dequeue(
-        &mut self,
-        latency: gst::ClockTime,
-        threshold: gst::ClockTime,
-        lookahead: gst::ClockTime,
-    ) -> Option<Vec<TranscriptItem>> {
-        let first_pts = self.items.front()?.pts;
-        if first_pts + latency > threshold {
-            // First item is too early to be sent to translation now
-            // we can wait for more items to accumulate.
-            return None;
-        }
-
-        // Can't wait any longer to send the first item to translation
-        // Try to get up to lookahead worth of items to improve translation accuracy
-        let limit = first_pts + lookahead;
-
-        let mut items_acc = vec![self.items.pop_front().unwrap()];
-        while let Some(item) = self.items.front() {
-            if item.pts > limit {
-                break;
-            }
-
-            items_acc.push(self.items.pop_front().unwrap());
-        }
-
-        Some(items_acc)
-    }
-}
-
 pub struct TranslateLoop {
     pad: glib::subclass::ObjectImplRef<TranslateSrcPad>,
     client: aws_translate::Client,
     input_lang: String,
     output_lang: String,
     tokenization_method: TranslationTokenizationMethod,
-    transcript_rx: mpsc::Receiver<Vec<TranscriptItem>>,
+    transcript_rx: mpsc::Receiver<Arc<Vec<TranscriptItem>>>,
     translate_tx: mpsc::Sender<Vec<TranslatedItem>>,
 }
 
@@ -121,7 +57,7 @@ impl TranslateLoop {
         input_lang: &str,
         output_lang: &str,
         tokenization_method: TranslationTokenizationMethod,
-        transcript_rx: mpsc::Receiver<Vec<TranscriptItem>>,
+        transcript_rx: mpsc::Receiver<Arc<Vec<TranscriptItem>>>,
         translate_tx: mpsc::Sender<Vec<TranslatedItem>>,
     ) -> Self {
         let aws_config = imp.aws_config.lock().unwrap();
@@ -175,12 +111,12 @@ impl TranslateLoop {
 
             let (ts_duration_list, content): (Vec<(gst::ClockTime, gst::ClockTime)>, String) =
                 transcript_items
-                    .into_iter()
+                    .iter()
                     .map(|item| {
                         (
                             (item.pts, item.duration),
                             match self.tokenization_method {
-                                Tokenization::None => item.content,
+                                Tokenization::None => item.content.clone(),
                                 Tokenization::SpanBased => {
                                     format!("{SPAN_START}{}{SPAN_END}", item.content)
                                 }