Improve language detection when using clip_timestamps

SYSTRAN · Jun 4, 2024 · 1becddb · 1becddb
1 parent 65551c0
commit 1becddb
Showing 1 changed file with 12 additions and 3 deletions.
diff --git a/faster_whisper/transcribe.py b/faster_whisper/transcribe.py
@@ -370,15 +370,24 @@ def transcribe(
                     or language_detection_segments < 1
                 ):
                     language_detection_segments = 1
-                seek = 0
+                if isinstance(clip_timestamps, str):
+                    start_timestamp = float(clip_timestamps.split(",")[0])
+                else:
+                    start_timestamp = clip_timestamps[0]
                 detected_language_info = {}
+                seek = int(
+                    start_timestamp * self.feature_extractor.sampling_rate
+                    // self.feature_extractor.hop_length
+                )
                 content_frames = (
                     features.shape[-1] - self.feature_extractor.nb_max_frames
                 )
+                end_frames = (
+                    seek + self.feature_extractor.nb_max_frames * language_detection_segments
+                )
                 while (
                     seek <= content_frames
-                    and seek
-                    < self.feature_extractor.nb_max_frames * language_detection_segments
+                    and seek < end_frames
                 ):
                     segment = features[
                         :, seek : seek + self.feature_extractor.nb_max_frames