OpenAI voice model detective story

detective-story.py

import asyncio

from openai import AsyncOpenAI
from openai.helpers import LocalAudioPlayer

import wave
import numpy as np

openai = AsyncOpenAI()

narrator_instructions = """
The speaker is narrating a detective story.

Voice Affect: Quietly dramatic.

Tone: Erudite. Learned. British accent.

Pacing: Slow and steady.

Emotion: Steady.

Pronunciation: Clear and precise.\n\nPauses: Brief pauses for narrative impact.
"""


def character_instructions(append):
    return f"""An American woman. A character in a detective story.

Speaking quickly. Trying not to be overheard.

Tone: {append}"""


narrator = [
    (
        "It was a dark and stormy night. The detective crept over to the window. A woman was talking to someone near the rosebeds.",
        narrator_instructions,
    ),
    (
        "Just then, the clock struck midnight. The woman turned, startled, and seemed to look right into the darkened room at the detective.",
        narrator_instructions,
    ),
    ("... she said, and then hurried around the corner.", narrator_instructions),
]

character = [
    ("shhh, we need to be quiet ...", character_instructions("Whispering.")),
    ("I don't think we have a chance of getting away with it.", character_instructions("Fearful.")),
    ("Is there someone there?", character_instructions("Startled.")),
]


async def main() -> None:
    player = LocalAudioPlayer()
    all_segments: list[np.ndarray] = []  # store audio buffers for wav output

    # Iterate over character and narrator tuples simultaneously
    for char_line, narr_line in zip(character, narrator):
        # Play the character line first
        char_input, char_instructions = char_line
        async with openai.audio.speech.with_streaming_response.create(
            model="gpt-4o-mini-tts",
            voice="alloy",  # character voice
            input=char_input,
            instructions=char_instructions,
            response_format="pcm",
        ) as response:
            # Convert streamed response to numpy buffer once
            buffer = await player._tts_response_to_buffer(response)
            await player.play(buffer)
            all_segments.append(buffer)

        # Then play the narrator line
        narr_input, narr_instructions = narr_line
        async with openai.audio.speech.with_streaming_response.create(
            model="gpt-4o-mini-tts",
            voice="ballad",  # narrator voice
            input=narr_input,
            instructions=narr_instructions,
            response_format="pcm",
        ) as response:
            buffer = await player._tts_response_to_buffer(response)
            await player.play(buffer)
            all_segments.append(buffer)

    # After all segments are played, save concatenated output to WAV
    if all_segments:
        combined = np.concatenate(all_segments, axis=0).flatten()
        int16_samples = (combined * 32767.0).astype(np.int16)
        with wave.open("output.wav", "wb") as wf:
            wf.setnchannels(1)  # mono
            wf.setsampwidth(2)  # 16-bit
            wf.setframerate(24000)  # match SAMPLE_RATE used by LocalAudioPlayer
            wf.writeframes(int16_samples.tobytes())


if __name__ == "__main__":
    asyncio.run(main())