Termux-STT

디바이스 리소스를 활용한 통합 온디바이스 음성인식(STT) 및 순수 Python 128d X-Vector 화자 분리 프레임워크
Unified On-Device Speech-to-Text Utilizing Device Resources & Pure Python 128d X-Vector Speaker Diarization

Architecture & Overview

Whisper.cpp, Vosk, Sherpa-ONNX 3대 네이티브 엔진을 통합하고, 닫힌 형태 순수 Python 128차원 X-Vector 클러스터링을 결합하여 80MB 미만의 초경량 메모리로 100% 온디바이스 실시간 음성인식과 화자 분리를 구현합니다.

Integrates Whisper.cpp, Vosk, and Sherpa-ONNX with a closed-form pure-Python 128-dimensional X-Vector clustering algorithm that operates in under 80MB RAM with zero cloud egress.

Asymmetric Hybrid Architecture Benchmark (Galaxy S25 - Snapdragon 8 Elite)

Test Audio: JFK Inaugural Address (60.59s mono 16kHz WAV)
Silicon: Qualcomm Snapdragon 8 Elite (Adreno 830 GPU + Oryon CPU)

Model	Pure CPU (4T)	Pure GPU (Vulkan)	Hybrid GPU-CPU (4T)	Hybrid Efficiency (1T)	Speedup / Advantage
Whisper Tiny	4.59s	9.56s	4.38s	5.48s	+4.6% vs CPU (Dispatch overhead minimized)
Whisper Small	15.48s	16.36s	12.93s	23.02s	+21.0% vs GPU, +16.5% vs CPU
Large-v3-Turbo	115.05s	87.87s	86.52s	90.39s	-28.5s vs CPU, CPU Load 0% during 80s encode

Multi-SoC Verification: Tested across Qualcomm Adreno 830, ARM Mali-G78 (Galaxy S21), and ARM Mali-G77 (Galaxy S20).
Upstream Contribution: Official PR submitted to ggml-org/whisper.cpp (PR #4089).

Installation & Quickstart

Python (PyPI)

pip install termux-stt

from termux_stt import create_engine

# 1. Initialize Engine with Hybrid GPU-Encoder / CPU-Decoder Acceleration
engine = create_engine("whisper", model="small", lang="en", threads=4, split_mode=True)

# 2. Transcribe Audio directly into Subtitles
result = engine.transcribe("samples/jfk_1min.wav")
print("Transcript:\n", result.text)
print("SRT Subtitles:\n", result.to_srt())

# 3. 2-Speaker Diarization without PyTorch
hybrid = create_engine("hybrid", lang="en", num_speakers=2)
diar_result = hybrid.diarize("samples/jfk_1min.wav")
for seg in diar_result.segments:
    print(f"[{seg.speaker}] ({seg.start:.1f}s -> {seg.end:.1f}s): {seg.text}")

CLI Command Line Usage

# 1. High-Performance Hybrid Transcribe (Vulkan GPU Encoder + 4 CPU Threads Decoder)
termux-stt transcribe samples/jfk_1min.wav --model small --split-mode

# 2. Ultra-Low Power Background Transcribe (GPU Encoder + 1 CPU Thread Decoder)
termux-stt transcribe meeting.wav --model small --optimize-1

# 3. Force Pure CPU Execution
termux-stt transcribe samples/jfk_1min.wav --device cpu --threads 4

Node.js / TypeScript (npm)

npm install termux-stt

const { createEngine } = require("termux-stt");

async function main() {
  // 1. Initialize Whisper Engine
  const engine = createEngine("whisper", { model: "tiny", lang: "en", threads: 4 });

  // 2. Transcribe Audio
  const result = await engine.transcribe("samples/jfk_1min.wav");
  console.log("Transcript:", result.text);
  console.log("SRT Subtitles:\n", result.toSrt());
}
main();

Official Documentation & Benchmarks

License

termux stt speech-to-text whisper whisper-cpp vosk sherpa-onnx diarization speaker-diarization voice-recognition audio-transcription on-device-ai edge-ai mobile-ai vulkan vulkan-compute gpu-acceleration real-time-factor low-latency arm64 android snapdragon adreno exynos arm-mali zero-compilation vad voice-activity-detection silero-vad x-vector k-means clustering srt-export vtt-export rttm microphone-streaming headless-audio pulseaudio offline-speech privacy-first termux-aichain termux-tts termux-llamacpp termux-diffusion ameva-runtime ggml quantization bionic-libc autonomous-agents voice-assistant

Metadata

Release files for termux-stt 1.3.0

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for termux-stt 1.3.0
File	Size	Uploaded
termux_stt-1.3.0.tar.gz	1.7 MB	Sep 28, 2026	Details

Built distribution (wheel)

Table of built distributions (wheels) for termux-stt 1.3.0
File	Interpreter	ABI	Platform	Reset
termux_stt-1.3.0-py3-none-any.whl 78.3 kB Sep 28, 2026	Python 3	none	any	Details

Total release size: 1.8 MB

Release files / termux_stt-1.3.0.tar.gz

Download URL	termux_stt-1.3.0.tar.gz
Size	1.7 MB
Tags	Source
SHA-256 checksum How to use checksums	`483db6202fc86d97b6cb33e640b261573bcd17b2148d4a1ebdd7efc8b051133a`
BLAKE2b-256 checksum How to use checksums	`4076292e0b540aaeef1498bd2f4596619fc587d258f660d3300f9753a50c9820`
Upload date	Sep 28, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/7.0.0 CPython/3.11.16`

Release files / termux_stt-1.3.0-py3-none-any.whl

Download URL	termux_stt-1.3.0-py3-none-any.whl
Size	78.3 kB
Tags	Python 3
SHA-256 checksum How to use checksums	`377a5bec5343fcb7a345effa55af2f3500eb99cc60aa3cbfe97758bd6187eaea`
BLAKE2b-256 checksum How to use checksums	`db260a6e0a171fccffaa88267a04146c824f6e79857bb1c725740e6b231a4f53`
Upload date	Sep 28, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/7.0.0 CPython/3.11.16`