RealtimeSTT-mac

A lightning-fast Apple Silicon-optimised Speech-to-Text & Voice-Activity-Detection library

---

About This Fork

This repository is a macOS-only, Apple Silicon-only fork of the excellent RealtimeSTT project by Kolja Beigel.

What Changed ❓

Area	Upstream	This fork
Target HW	cross-platform (CPU & CUDA)	Apple Silicon only
Backend	faster-whisper + CTranslate2 (generic)	faster-whisper + CTranslate2 built with Apple Accelerate
Packaging	setup.py, requirements.txt	PEP 621 (pyproject.toml) + uv
Porcupine	pvporcupine==1.9.5 did not support Apple Silicon	pvporcupine>=3.0.0 does support Apple Silicon!
Packaging	setup.py, requirements.txt	PEP 621 (pyproject.toml) + uv
Server / WebSocket	optional STT server & client	Removed – local library only
Docker images	CUDA / CPU images	Removed
Wheels	fetched from PyPI	Bundled fat-wheels (contain the Accelerate dylib)

Big thanks to Kolja Beigel for the original codebase – all core algorithms, examples and documentation originate from his work. This fork simply re-targets the library for M-series Macs and trims anything that is not useful on that platform.

---

Features

• Realtime VAD
  • WebRTC-VAD (fast)
  • Optional Silero VAD (accurate, runs via PyTorch)
• Realtime & Batch Transcription Powered by faster-whisper, executed with CTranslate2 + Apple Accelerate
• Wake-Word Detection
  • Picovoice Porcupine (pvporcupine)
  • OpenWakeWord (openwakeword)
• Low-latency audio pipeline < 200 ms from speech to text on an M2 Pro
• Pythonic callbacks for start/stop, partial results, wake-word events
• Pure Python install – no Xcode project, no CMake build required

---

Installation (uv)

Requires macOS 12+ and M-series CPU (M1/M2/M3).

# Initialise a new uv project (if you haven't already)
uv init -p 3.11            # or -p 3.12

# Add RealtimeSTT-mac from Git or PyPI
uv add realtimestt-mac

# Activate & run
uv run python - <<'PY'
from RealtimeSTT import AudioToTextRecorder

rec = AudioToTextRecorder(device="cpu")  # Apple Accelerate is used automatically
print("Speak now…")
while True:
    rec.text(print)
PY

The wheels directory is bundled inside the package and contains pre-built, self-contained CTranslate2 binaries for:

• cp311-macosx_arm64
• cp312-macosx_arm64

No CUDA, x86_64 or Windows/Linux wheels are shipped.

---

Quick Example

from RealtimeSTT import AudioToTextRecorder

def on_text(txt):
    print("📝", txt)

with AudioToTextRecorder(
        device="cpu",                   # always "cpu" on Apple Silicon
        enable_realtime_transcription=True,
        wakeword_backend="pvporcupine", # say "Hey Siri" etc.
) as recorder:
    print("Speak or say your wake-word…")
    while True:
        recorder.text(on_text)

---

Removing Server / Docker Components

The old WebSocket server, CLI client, Dockerfiles and GPU requirements have been removed. If you need a networked STT service you can still run the upstream project.

---

Dependencies

Package	Why it stays
PyAudio	low-latency microphone input
torch / torchaudio	Silero VAD & OpenWakeWord
faster-whisper	Whisper inference
ctranslate2	Execution backend (Accelerate build)
pvporcupine, openwakeword	Wake-word detection
scipy, soundfile, webrtcvad-wheels	signal processing

Removed: websockets, websocket-client, halo (spinner) and all CUDA-specific dependencies.

---

Limitations

• Apple Silicon only – will not install on Intel Macs or other OSes
• CPU backend only – Accelerate uses CPU + NEON; GPU (Metal) is not used
• No CUDA / ROCm – use upstream for NVIDIA/AMD GPUs
• No server mode – library-only usage

---

Credits

• Original code ⊕ algorithm design: Kolja Beigel https://github.com/KoljaB/RealtimeSTT
• Apple Silicon port & maintenance: Tom Whiting https://github.com/tomWhiting/RealtimeSTT-mac

Licensed under the MIT License (see LICENSE file).

Enjoy ultra-fast transcription on your M-series Mac! PRs and feedback welcome.