From 6f02a2f2a90a3f884a44735a6b58aa75a4fea4c2 Mon Sep 17 00:00:00 2001
From: Sofia Casadei <69089170+sofi444@users.noreply.github.com>
Date: Tue, 27 May 2025 20:54:33 +0200
Subject: [PATCH] chunk speech after s if no pause detected by VAD (#328)

* chunk speech after s if no pause detected by VAD

* add attr descriptions in AlgoOptions

* Fix

---------

Co-authored-by: Freddy Boulton <41651716+freddyaboulton@users.noreply.github.com>
---
 backend/fastrtc/reply_on_pause.py | 18 +++++++++++++++++-
 1 file changed, 17 insertions(+), 1 deletion(-)

diff --git a/backend/fastrtc/reply_on_pause.py b/backend/fastrtc/reply_on_pause.py
index c4c6ba7..cd47a98 100644
--- a/backend/fastrtc/reply_on_pause.py
+++ b/backend/fastrtc/reply_on_pause.py
@@ -18,11 +18,20 @@ logger = getLogger(__name__)
 
 @dataclass
 class AlgoOptions:
-    """Algorithm options."""
+    """
+    Algorithm options.
+
+    Attributes:
+    - audio_chunk_duration: Duration in seconds of audio chunks passed to the VAD model.
+    - started_talking_threshold: If the chunk has more than started_talking_threshold seconds of speech, the user started talking.
+    - speech_threshold: If, after the user started speaking, there is a chunk with less than speech_threshold seconds of speech, the user stopped speaking.
+    - max_continuous_speech_s: Max duration of speech chunks before the handler is triggered, even if a pause is not detected by the VAD model.
+    """
 
     audio_chunk_duration: float = 0.6
     started_talking_threshold: float = 0.2
     speech_threshold: float = 0.1
+    max_continuous_speech_s: float = float("inf")
 
 
 @dataclass
@@ -216,7 +225,14 @@ class ReplyOnPause(StreamHandler):
                     state.stream = audio
                 else:
                     state.stream = np.concatenate((state.stream, audio))
+
+                # Check if continuous speech limit has been reached
+                current_duration = len(state.stream) / sampling_rate
+                if current_duration >= self.algo_options.max_continuous_speech_s:
+                    return True
             state.buffer = None
+
+            # Check if a pause has been detected by the VAD model
             if dur_vad < self.algo_options.speech_threshold and state.started_talking:
                 return True
         return False