fallback on whisper alignment failures, update readme

README.md

@@ -23,7 +23,6 @@
   <a href="#setup">Setup</a> •
   <a href="#example">Usage</a> •
   <a href="#other-languages">Multilingual</a> •
-  <a href="#python-usage">Python</a> •
   <a href="#contribute">Contribute</a> •
   <a href="EXAMPLES.md">More examples</a>
 </p>
@@ -33,7 +32,6 @@
 <img width="1216" align="center" alt="whisperx-arch" src="https://user-images.githubusercontent.com/36994049/208313881-903ab3ea-4932-45fd-b3dc-70876cddaaa2.png">
 
 
-
 <p align="left">Whisper-Based Automatic Speech Recognition (ASR) with improved timestamp accuracy using forced alignment.
 
 </p>
@@ -55,6 +53,20 @@ Install this package using
 
 `pip install git+https://github.com/m-bain/whisperx.git`
 
+If already installed, update package to most recent commit
+
+`pip install git+https://github.com/m-bain/whisperx.git --upgrade`
+
+If wishing to modify this package, clone and install in editable mode:
+```
+$ git clone https://github.com/m-bain/whisperX.git
+$ cd whisperX
+$ pip install -e .
+```
+
+
+`pip install git+https://github.com/m-bain/whisperx.git --upgrade`
+
 You may also need to install ffmpeg, rust etc. Follow openAI instructions here https://github.com/openai/whisper#setup.
 
 <h2 align="left" id="example">Usage 💬 (command line)</h2>

requirements.txt

@@ -2,6 +2,7 @@ numpy
 torch
 torchaudio
 tqdm
+soundfile
 more-itertools
 transformers>=4.19.0
 ffmpeg-python==0.2.0

whisperx/alignment.py

@@ -64,7 +64,8 @@ def backtrack(trellis, emission, tokens, blank_id=0):
             if j == 0:
                 break
     else:
-        raise ValueError("Failed to align")
+        # failed
+        return None
     return path[::-1]
 
 # Merge the labels

whisperx/transcribe.py

@@ -293,7 +293,11 @@ def align(
     word_segments_list = []
     for idx, segment in enumerate(transcript):
         if int(segment['start'] * SAMPLE_RATE) >= audio.shape[1]:
-            # original whisper error, transcript is outside of duration of audio, not possible. Skip to next (finish).
+            print("Failed to align segment: original start time longer than audio duration, skipping...")
+            continue
+        
+        if int(segment['start']) >= int(segment['end']):
+            print("Failed to align segment: original end time is not after start time, skipping...")
             continue
 
         t1 = max(segment['start'] - extend_duration, 0)
@@ -325,11 +329,16 @@ def align(
         t_words_nonempty_idx = [x for x in range(len(t_words_clean)) if t_words_clean[x] != ""]
         segment['word-level'] = []
 
+        fail_fallback = False
         if len(t_words_nonempty) > 0:
             transcription_cleaned = "|".join(t_words_nonempty).lower()
             tokens = [model_dictionary[c] for c in transcription_cleaned]
             trellis = get_trellis(emission, tokens)
             path = backtrack(trellis, emission, tokens)
+            if path is None:
+                print("Failed to align segment: backtrack failed, resorting to original...")
+                fail_fallback = True
+            else:
                 segments = merge_repeats(path, transcription_cleaned)
                 word_segments = merge_words(segments)
                 ratio = waveform_segment.size(0) / (trellis.size(0) - 1)
@@ -372,6 +381,9 @@ def align(
                         else:
                             word_segments_list[-1]['text'] += ' ' + curr_word
         else:
+            fail_fallback = True
+
+        if fail_fallback:
             # then we resort back to original whisper timestamps
             # segment['start] and segment['end'] are unchanged
             prev_t2 = 0