Whisper (Local)

The WhisperSTTService allows you to run OpenAI's Whisper models locally on your own hardware. It supports multiple backends, including Faster Whisper for general CPUs/GPUs and MLX Whisper specifically optimized for Apple Silicon.

Installation

Depending on your hardware, choose the appropriate installation:

For General Hardware (Faster Whisper)

pip install "piopiy-ai[whisper]"

For Apple Silicon (MLX)

pip install "piopiy-ai[mlx-whisper]"

Configuration

WhisperSTTService (General) Parameters

Parameter	Type	Default	Description
model	str | Model	DISTIL_MEDIUM_EN	Model size (e.g., base, medium, large-v3).
device	str	"auto"	Device to run on (cpu, cuda, auto).
compute_type	str	"default"	Precision (int8, float16, etc.).
language	Language	EN	Transcription language.

WhisperSTTServiceMLX (Apple Silicon) Parameters

Parameter	Type	Default	Description
model	str | MLXModel	TINY	MLX model repository/ID.
temperature	float	0.0	Sampling temperature.

Usage

Basic Setup (Faster Whisper)

from piopiy.services.whisper.stt import WhisperSTTService, Model
from piopiy.transcriptions.language import Language

stt = WhisperSTTService(
    model=Model.BASE,
    device="cpu", # or "cuda" for NVIDIA GPUs
    language=Language.EN
)

Apple Silicon Optimized (MLX)

from piopiy.services.whisper.stt import WhisperSTTServiceMLX, MLXModel

stt = WhisperSTTServiceMLX(
    model=MLXModel.LARGE_V3_TURBO,
    language=Language.EN
)

Notes

• Initial Run: The first time you run a specific model, Piopiy will download several hundred megabytes (or gigabytes) of model weights from Hugging Face.
• Privacy: Local Whisper processing ensures that your audio data never leaves your server.
• HW Acceleration: For NVIDIA GPUs, ensure you have appropriate CUDA libraries installed to use device="cuda".