crop samples to be around 1 second

crop_audio_files.py

@@ -27,7 +27,7 @@ def copy_sample(data: Any, sample_rate: int, start: float, end: float):
 # [id, audio data, sample rate, extension, words]
 TSegment = Tuple[str, Any, int, str, str]
 
-def crop_audio_file(audio_file: str, ctm_file: str, target_duration_secs: float, max_word_duration_secs: float = 5.0, min_segment_duration_secs: float = 12.0, max_segment_duration_secs: float = 21.0) -> List[TSegment]:
+def crop_audio_file(audio_file: str, ctm_file: str, target_duration_secs: float, max_word_duration_secs: float = 2.0, min_segment_duration_secs: float = 0.5, max_segment_duration_secs: float = 1.5) -> List[TSegment]:
     data, sample_rate = sf.read(audio_file)
     ctm = read_ctm_file(ctm_file)
 
@@ -46,27 +46,29 @@ def crop_audio_file(audio_file: str, ctm_file: str, target_duration_secs: float,
         return (segment_id, audio_sample, sample_rate, ext, words)
 
     for _, _, word_start, word_duration_secs, word in ctm:
-        if cur_start is None:
-            cur_start = word_start
-
-        word_end = word_start + word_duration_secs
-
-        if cur_end - cur_start > target_duration_secs or word_end - cur_start >= max_segment_duration_secs or word_duration_secs > max_word_duration_secs:
-            if cur_end - cur_start >= min_segment_duration_secs:
+        # exclude pathological alignments
+        if word_duration_secs > max_word_duration_secs:
+            if cur_start is not None and cur_end - cur_start >= min_segment_duration_secs and cur_end - cur_start <= max_segment_duration_secs:
                 segment = make_segment(cur_start, cur_end, cur_words)
                 segments.append(segment)
+            cur_start = None
             cur_words = []
-            # exclude pathological alignments
-            if word_duration_secs > max_word_duration_secs:
-                cur_start = None
-                continue
-            cur_start = word_start
+            continue
 
+        if cur_start is None:
+            cur_start = word_start
 
-        cur_words.append(word)
+        word_end = word_start + word_duration_secs
         cur_end = word_end
+        cur_words.append(word)
+
+        if cur_end - cur_start >= target_duration_secs - 0.1:
+            segment = make_segment(cur_start, cur_end, cur_words)
+            segments.append(segment)
+            cur_start = None
+            cur_words = []
 
-    if cur_start is not None and cur_end - cur_start >= min_segment_duration_secs:
+    if cur_start is not None and cur_end - cur_start >= min_segment_duration_secs and cur_end - cur_start <= max_segment_duration_secs:
         segment = make_segment(cur_start, cur_end, cur_words)
         segments.append(segment)
 
@@ -105,7 +107,7 @@ if __name__ == "__main__":
     ctm_files_dir = 'data/ctm/words'
     output_audio_dir = 'data/audio_cropped'
     output_metadata_file = 'metadata_cropped.csv'
-    target_len_secs = 15.6
+    target_len_secs = 1
 
     n = crop_audio_files(input_audio_dir, ctm_files_dir, output_audio_dir, output_metadata_file, target_len_secs)
     print(f'{n} crops created')

data/audio_cropped.tar.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f891bd1f399edb9f2acd9701c301091ad5ebeed3aa50a4e273d34a6b8f030972
-size 795604076
+oid sha256:73801b326a6059a27269efffc9c981caee8f14e6e36c04d4f20668bc9a4fd2ec
+size 725365601