transcribe.py

# /// script
# requires-python = ">=3.13"
# dependencies = [
#     "audioop-lts",
#     "llm",
#     "llm-gemini",
#     "pyaudio",
#     "pydub",
# ]
# ///
import llm
from pydub import AudioSegment
import io
import json

def split_audio_into_chunks(audio_path, chunk_length_minutes=5):
    """
    Split an m4a audio file into N-minute chunks and return each chunk as binary data.
    """
    # Load the audio file
    audio = AudioSegment.from_file(audio_path, format="m4a")

    # Calculate chunk length in milliseconds
    chunk_length_ms = chunk_length_minutes * 60 * 1000

    # Initialize list for binary chunks
    binary_chunks = []

    # Calculate how many chunks we'll have
    total_length_ms = len(audio)

    # Split the audio into chunks
    for start_ms in range(0, total_length_ms, chunk_length_ms):
        end_ms = min(start_ms + chunk_length_ms, total_length_ms)
        chunk = audio[start_ms:end_ms]

        # Convert chunk to binary data in mp3 format
        buffer = io.BytesIO()
        chunk.export(buffer, format="mp3")
        binary_chunks.append(buffer.getvalue())

    return binary_chunks

# Example usage:
audio_path = "31.m4a"
model = llm.get_model("gemini-2.0-flash")
schema = llm.schema_dsl("timestamp str: mm:ss,text,speaker: SPK_0 or SPK_1,language: two letter code", multi=True)
prompt = "transcribe"
chunk_length_minutes = 5

chunks = split_audio_into_chunks(audio_path, chunk_length_minutes=chunk_length_minutes)
output_data = []

for i, chunk in enumerate(chunks):
    try:
        attachment = llm.Attachment(type="audio/mp3", content=chunk)
        response = model.prompt("transcribe", attachments=[attachment], schema=schema)

        # Parse JSON for the current chunk
        response_json = json.loads(response.text())["items"]

        # --- COMPUTE AND APPLY OFFSET ---
        # How many milliseconds in each chunk
        chunk_length_ms = chunk_length_minutes * 60 * 1000
        # The start of the current chunk (in ms)
        chunk_start_ms = i * chunk_length_ms

        # Convert that to total seconds for easy arithmetic
        chunk_start_seconds = chunk_start_ms // 1000

        for item in response_json:
            # item["timestamp"] is something like "mm:ss"
            mm_str, ss_str = item["timestamp"].split(":")
            mm = int(mm_str)
            ss = int(ss_str)
            original_seconds = mm * 60 + ss

            # Add the offset
            updated_seconds = original_seconds + chunk_start_seconds

            # Convert back to mm:ss
            new_mm = updated_seconds // 60
            new_ss = updated_seconds % 60
            item["timestamp"] = f"{new_mm:02d}:{new_ss:02d}"

        # Now that timestamps are offset, extend the main list
        output_data.extend(response_json)

    except Exception as e:
        print(f"Error: {e}")
        break

    # Optionally break early
    #if i >= 2:
    #    break

# Finally, write out to JSON
with open("31.json", 'w', encoding='utf-8') as file:
    json.dump(output_data, file, indent=2, ensure_ascii=False)