How can I detect and correct repeated phrases in Whisper transcriptions?

Below is one way to adapt your approach. Instead of looking only for single-word repeats, you can slide over the text checking for repeated sequences (or “n-grams”) of different lengths. For example, the function below checks for repetitions for phrase lengths from 1 up to a maximum (which you can tune):

PYTHONCopied
def detect_repeated_phrases(text, max_phrase_length=5):
    """
    Detects repeated phrases in a text.
    A "repeated phrase" refers to a sequence of words that immediately repeats.
    For example, in "hello hello world", the phrase "hello" repeats,
    and in "this is a test this is a test", the phrase "this is a test" repeats.
    
    Parameters:
        text (str): Input transcription text.
        max_phrase_length (int): Maximum number of words to consider as a phrase.
        
    Returns:
        A set of repeated phrases detected in the text.
    """
    words = text.split()
    repeated_phrases = set()
    n = len(words)
    
    # For each possible phrase length from 1 to max_phrase_length
    for phrase_length in range(1, max_phrase_length + 1):
        # Slide over the words such that there is room for two consecutive groups
        for i in range(n - phrase_length * 2 + 1):
            first_phrase = words[i:i + phrase_length]
            second_phrase = words[i + phrase_length:i + 2 * phrase_length]
            
            # If both phrase groups are identical, record the phrase.
            if first_phrase == second_phrase:
                repeated_phrases.add(" ".join(first_phrase))
                
                # Optionally, you can skip ahead to avoid overlapping detection:
                # i += phrase_length  # Uncomment if you want to avoid re-checking overlapping phrases.
                
    return repeated_phrases

# Example usage:
if __name__ == "__main__":
    input_text = "this is a test this is a test and a hello hello world"
    found_phrases = detect_repeated_phrases(input_text)
    print("Detected repeated phrases:", found_phrases)

Explanation

• Sliding Window:
  We iterate over every possible position in the text that can hold two consecutive groups of size phrase_length. This lets us check for immediate repetition.

• Phrase Length Flexibility:
  By varying the phrase_length from 1 (single words) up to max_phrase_length, you can catch both single-word repeats and longer phrases.

• Output:
  The function returns a set of unique repeated phrases. You could also adjust the function to return additional information (such as counts or positions) if needed.

Possible Improvements

• Fuzzy Matching:
  If you want to catch repetitions where one instance might have a minor error (for example, one repeated sequence with an extra punctuation), you could incorporate fuzzy string matching or preprocessing (like lowercasing and stripping punctuation).

• Overlapping Matches:
  Depending on your needs, you might want to avoid detecting overlapping repeated phrases. In that case, once you find a repeated phrase, you can advance the index to skip ahead accordingly.

This approach should be a more intelligent way to detect consecutive repeated phrases of varying lengths in your transcription output.