Skip to content

Instantly share code, notes, and snippets.

Created March 27, 2023 02:33
Show Gist options
  • Save endolith/5f539047b4b6f3f9a9549afa3e2f9ad5 to your computer and use it in GitHub Desktop.
Save endolith/5f539047b4b6f3f9a9549afa3e2f9ad5 to your computer and use it in GitHub Desktop.
Transcribe a long audio recording using OpenAI Whisper API
Break up a long recording to fit within the Whisper API's limits, with some
overlap, so no words are missed, and then feed to OpenAI Whisper API to
transcribe it to .txt file. Written by endolith and ChatGPT-4.
import openai
import math
import os
import subprocess
openai.api_key = 'sk-YOUR_API_KEY_HERE'
filename = r'C:/Users/YOUR/PATH/FILE.m4a'
# Constants
max_bytes = 26214400 # From Whisper error message
overlap_seconds = 5
# Get the bit rate directly from the file
bit_rate = float(subprocess.check_output(
["ffprobe", "-v", "quiet", "-show_entries", "format=bit_rate", "-of",
"default=noprint_wrappers=1:nokey=1", filename]).strip())
# Estimate the duration of each chunk
chunk_duration_s = (max_bytes * 8.0) / bit_rate * 0.9
# Get the duration of the audio file
audio_duration_s = float(subprocess.check_output(
["ffprobe", "-v", "quiet", "-show_entries", "format=duration", "-of",
"default=noprint_wrappers=1:nokey=1", filename]).strip())
# Calculate the number of chunks
num_chunks = math.ceil(audio_duration_s / (chunk_duration_s - overlap_seconds))
transcriptions = []
output_folder = "chunks"
os.makedirs(output_folder, exist_ok=True)
# Get the file extension from the filename
file_extension = os.path.splitext(filename)[1]
for i in range(num_chunks):
start_s = i * (chunk_duration_s - overlap_seconds)
end_s = start_s + chunk_duration_s
# Save the chunk to disk
chunk_file = os.path.join(output_folder, f"chunk_{i + 1}{file_extension}")
# Use ffmpeg to extract the chunk directly into the compressed format (m4a)["ffmpeg", "-ss", str(start_s), "-i", filename, "-t",
str(chunk_duration_s), "-vn", "-acodec", "copy", "-y",
# Transcribe the chunk
with open(chunk_file, "rb") as file:
transcription = openai.Audio.transcribe("whisper-1", file)
# Save transcriptions to a file
with open("transcriptions.txt", "w") as file:
for idx, transcription in enumerate(transcriptions):
file.write(f"Chunk {idx + 1}:\n{transcription}\n\n")
Copy link

rafapg commented Dec 19, 2023

To support the new version of whisper API, the line 57 should be changed to

        transcription =

and a client initialization should be done at the start of the file:

openai.api_key =  "[...]"
client = openai.OpenAI()

Copy link

You should be passing the previous chunk as a prompt... this wag you can avoid accidently recording the same word twice.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment