How to handle large audio files with Whisper API in openai-python without hitting memory limits?

[Super-Mutec17]

Hi,
I’m working with Whisper API in openai-python to transcribe long recordings (over 1 hour), and I’m running into memory and timeout issues when sending large files in one request.

[danowar2]

Think I need anything from the center

[SameerSenapati17]

If you’re running into memory or timeout issues with Whisper API when transcribing long recordings (1+ hours), it’s best to avoid sending the entire file in a single request. Large uploads increase both memory usage and the chance of hitting request timeouts.

[manimovassagh]

The Whisper API has a 25MB file size limit per request. For long recordings (1+ hour), you need to split the audio into chunks before sending. Here's a reliable approach using pydub:

from pydub import AudioSegment
from openai import OpenAI
import io

client = OpenAI()

def transcribe_large_file(file_path: str, chunk_length_ms: int = 10 * 60 * 1000) -> str:
    """Transcribe a large audio file by splitting into chunks.

    Args:
        file_path: Path to the audio file
        chunk_length_ms: Chunk size in milliseconds (default 10 minutes)
    """
    audio = AudioSegment.from_file(file_path)

    full_transcript = []

    for i in range(0, len(audio), chunk_length_ms):
        chunk = audio[i:i + chunk_length_ms]

        # Export chunk to buffer (avoids writing temp files to disk)
        buf = io.BytesIO()
        chunk.export(buf, format="mp3", bitrate="64k")  # compress to stay under 25MB
        buf.seek(0)
        buf.name = "chunk.mp3"

        transcript = client.audio.transcriptions.create(
            model="whisper-1",
            file=buf,
            response_format="text",
        )

        full_transcript.append(transcript)
        print(f"Transcribed chunk {i // chunk_length_ms + 1}")

    return " ".join(full_transcript)

result = transcribe_large_file("long_recording.wav")

A few things to keep in mind:

• 10-minute chunks at 64kbps MP3 stay well under the 25MB limit while maintaining good transcription quality
• Overlap chunks by a few seconds if you're getting words cut off at boundaries -- split at silence points when possible using pydub.silence.detect_silence()
• Use response_format="verbose_json" if you need timestamps, then stitch them together by adding offsets
• For memory efficiency, avoid loading the entire file with AudioSegment if it's extremely large. Use ffmpeg directly to split:

# Split a file into 10-minute chunks without loading into memory
ffmpeg -i long_recording.wav -f segment -segment_time 600 -c:a libmp3lame -b:a 64k chunk_%03d.mp3

Then iterate over the chunk files and transcribe each one.