Split large audio file and transcribe it using the Whisper API from OpenAI

transcribe.py

import os
import sys
import openai
import os.path

from dotenv import load_dotenv
from pydub import AudioSegment

load_dotenv()
openai.api_key = os.getenv('OPENAI_API_KEY')

audio = AudioSegment.from_mp3(sys.argv[1])

segment_length = 25 * 60
duration = audio.duration_seconds
print('Segment length: %d seconds' % segment_length)
print('Duration: %d seconds' % duration)

segment_filename = os.path.basename(sys.argv[1])
segment_filename = os.path.splitext(segment_filename)[0]
number_of_segments = int(duration / segment_length)
segment_start = 0
segment_end = segment_length * 1000
enumerate = 1
prompt = ""

for i in range(number_of_segments):
    sound_export = audio[segment_start:segment_end]

    exported_file = '/tmp/' + segment_filename + '-' + str(enumerate) + '.mp3'
    sound_export.export(exported_file, format="mp3")
    print('Exported segment %d of %d' % (enumerate, number_of_segments))

    f = open(exported_file, "rb")
    data = openai.Audio.transcribe("whisper-1", f, prompt=prompt)
    f.close()
    print('Transcribed segment %d of %d' % (enumerate, number_of_segments))

    f = open(os.path.join('transcripts', segment_filename + '.txt'), "a")
    f.write(data.text)
    f.close()

    prompt += data.text
    segment_start += segment_length * 1000
    segment_end += segment_length * 1000
    enumerate += 1

Hi, thanks for sharing this code. Just as a warning, in line 21 you use int() to find the number of segments. This will lead to dropped endings, as it rounds down. e.g. you have a length of 3200sec, your code will determine it's 2 segments of 1500sec and drop the last 200sec. math.ceil() would be the correct function.
Also I believe your openAI API is not quite up to date anymore, I had to adjust it to:

from openai import OpenAI
client = OpenAI()
data = client.audio.transcriptions.create(model="whisper-1", file=f)

@ceinem

Here is a fix.

import math
from pydub import AudioSegment


MAX_SEGMENT_LENGTH: int = 10  # minutes


audio = AudioSegment.from_file('...some-audio-file.mp3')

# total number of segments 
n_seg: int = math.ceil(audio.duration_seconds / (MAX_SEGMENT_LENGTH * 60))

# duration per segment in seconds. Consistent for all segments.
seg_duration: float = audio.duration_seconds / n_seg

# can splice audio 
segs: list[AudioSegment] = [
    audio[i * seg_duration * 1000:(i + 1) * seg_duration * 1000]
    for i in range(n_seg)
]

# can now concurrently transcribe with OpenAI 

I'm curious why you're adding the previous segment of transcription into the prompt for future segment transcriptions here? Docs from OpenAI says that prompt ignores anything over 224 tokens.

In addition, the prompt is limited to only 224 tokens. If the prompt is longer than 224 tokens, only the final 224 tokens of the prompt will be considered; all prior tokens will be silently ignored.