How do I update Whisper API usage for long videos and correctly integrate batch timestamps?

Below are the answers to your questions:

1. Using Whisper v3 Through the API

Currently, there isn’t a Whisper v3 model available via the API. The model identifier “whisper-1” is used even for the latest version of Whisper provided through our API. In other words, there’s no need—and no option—to explicitly specify “v3”; the API always uses the up-to-date implementation under the “whisper-1” name.

2. Free Instances vs. API Pricing

The free versions of Whisper you see online are the open-source releases. When using the API, you’re accessing a hosted, optimized inference process that offers improved performance and scalability. The API pricing covers the cost of providing this managed infrastructure and additional efficiencies that aren’t included in the open-source version.

3. File Size Limitation and Timestamp Integration

Yes, the current API version has a file size limit of 25MB per request. For a 40-minute file exceeding 500MB, you will need to split the audio into smaller segments.

When splitting the file into batches, you must handle timestamp integration manually. A common approach is to keep track of each segment’s start time (an offset) relative to the full audio. After obtaining the transcription (or translation) for each batch, adjust the timestamps by adding the offset. This way, when you merge the results into a single SRT file, the timestamps correctly reflect the continuous time of the original audio.

Below is an example in Python for adjusting timestamps:

PYTHONCopied
def adjust_timestamps(segment_transcript, offset_seconds):
    """
    Adjust the SRT file timestamps for a segment by adding the given offset.
    
    segment_transcript: List of subtitles, each with start and end times as seconds.
    offset_seconds: The time offset in seconds for this segment.
    """
    adjusted = []
    for subtitle in segment_transcript:
        adjusted_subtitle = {
            'start': subtitle['start'] + offset_seconds,
            'end': subtitle['end'] + offset_seconds,
            'text': subtitle['text']
        }
        adjusted.append(adjusted_subtitle)
    return adjusted

# Example usage:
# segment_transcript = [{'start': 0.0, 'end': 5.0, 'text': 'Hello world'}, ...]
# offset_seconds = 60  # For a segment starting at 1 minute of the original audio
# adjusted_result = adjust_timestamps(segment_transcript, offset_seconds)

This approach tracks each segment’s timing so that the final merged SRT file aligns with the original full-length audio.

Let me know if you need more details on any of these points!