Termux-STT: Enterprise On-Device Speech-to-Text & Speaker Diarization

Termux-STT is an industrial-grade, zero-compilation on-device Speech-to-Text (STT) and multi-speaker diarization framework engineered specifically for Android Termux, ARM64 mobile hardware, and edge environments. By orchestrating a Tri-Engine acoustic pipeline (Whisper.cpp, Vosk/Kaldi, and Sherpa-ONNX Zipformer) with pure-Python x-vector speaker clustering and direct Vulkan GPU acceleration, Termux-STT achieves sub-realtime transcription speeds (up to 15.7x faster than real-time, RTF 0.064x) and continuous offline listening with zero cloud telemetry.

1. Installation Guide

Termux-STT is distributed across both Python (PyPI) and Node.js (npm) ecosystems. It runs in unprivileged user-space on Android Termux (ARM64) and Linux aarch64/x86_64.

1.1 Prerequisites on Android Termux

Update package repositories and install foundational audio and build utilities:

pkg update -y
pkg install -y clang python python-numpy nodejs termux-api ffmpeg pulseaudio

1.2 Pure Package Installation (Zero-Compilation Bundled Binary)

Install the core package from PyPI via pip:

pip install --upgrade pip
pip install termux-stt

To install with development extras:

pip install "termux-stt[dev]"

1.3 Node.js / TypeScript SDK & CLI Installation

Install globally or locally via npm:

# Global CLI installation (bridges to underlying Python runtime)
npm install -g termux-stt

# Project dependency installation
npm install termux-stt

1.4 Post-Installation Automated Environment Provisioner (termux-stt install)

While pure package installation provides instant offline WAV inference, production deployments involving compressed media (MP3/M4A/FLAC), live microphone streaming, or mobile GPU acceleration require full environment provisioning. Run the automated 1-click provisioner:

termux-stt-install
# Or equivalently:
termux-stt install

The automated engine installer (EngineInstaller) executes a 3-stage provisioning pipeline:

1. Native System Dependencies (install_system_dependencies): Automatically invokes Termux pkg to install ffmpeg, libbluray, libxml2, git, termux-api, and curl, enabling universal audio decoding and microphone capture via Android APIs.
2. Adaptive Engine Binary Provisioning (install_whisper_cpp):
   • Fast-Track Stream Extractor (~3s): On Android ARM64 Termux, precompiled Vulkan+NEON Bionic binaries (whisper-cli-android-arm64.tar.gz) are automatically extracted from GitHub Releases in ~3 seconds, completely eliminating 20-minute on-device compilation and mobile OOM aborts.
   • Zero-Hardcoding SSOT Endpoints: Binary downloads dynamically route through unified SSOT candidate endpoints (TERMUX_STT_RELEASE_TAG -> v{__version__} -> releases/latest/download -> uno-km/ameva-runtime releases fallback) with automated fallback to on-device C++ compilation (cmake + clang) in offline or air-gapped environments.
   • Bundled Package Binary Fallback: Automatically discovers and links bundled termux_stt/bin/whisper-cli if present.
3. Sub-Engine Ecosystem Provisioning (install_vosk, install_sherpa_onnx): Provisions vosk for sub-30ms real-time streaming and sherpa-onnx for next-generation ONNX Zipformer models, while pre-initializing model cache structures in ~/.cache/termux-stt/models/.

1.5 Pure Install vs. Post-Install Comparison Matrix

2. GPU Hardware Acceleration Provisioning (ameva-runtime)

To unlock mobile GPU tensor compute via Vulkan SPIR-V compute pipelines on Qualcomm Adreno or ARM Mali silicon, pair termux-stt with the unified @ameva/runtime hardware acceleration layer.

2.1 Unified Installation Command

Install both the STT engine and the hardware acceleration runtime simultaneously:

# Python Environment
pip install termux-stt ameva-runtime

# Node.js / JavaScript Environment
npm install -g termux-stt @ameva/runtime

2.2 Hardware Diagnostics & Zero-Silent-Fallback Protocol

Verify Vulkan driver detection and SIMD feature availability:

termux-stt doctor

Termux-STT strictly enforces a Zero-Silent-Fallback Protocol:

• When --device vulkan or --device gpu is requested and ameva-runtime is not installed, the engine immediately halts with [ERROR: AMEVA-STT-E001] rather than silently degrading to CPU execution.
• If no compatible Vulkan driver (/system/lib64/libvulkan.so) is found, the engine halts with [ERROR: AMEVA-STT-E002], preventing unexpected battery drain and thermal throttling.

3. Comprehensive Usage Guide & Manual

Termux-STT provides intuitive, high-performance interfaces across CLI, Python, and Node.js.

3.1 Command-Line Interface (CLI)

# 1. Standard High-Accuracy Transcription (Greedy Search default: -bs 1)
termux-stt transcribe meeting.wav -e whisper -m base -l en

# 2. Hardware-Accelerated Vulkan GPU Execution
termux-stt transcribe speech.wav -e whisper -m small -d vulkan

# 3. Multi-Beam Exploration for Precision Workloads (Explicit Beam Size Override)
termux-stt transcribe legal_deposition.wav -m small -bs 5

# 4. Voice Activity Detection (VAD) Pre-filtering (Drops Silence)
termux-stt transcribe lecture.wav -m base --vad

# 5. Export to Timestamped Subtitles (SRT / VTT / JSON)
termux-stt transcribe interview.mp3 --format srt -o output.srt

# 6. Audio Translation to English on the Fly
termux-stt transcribe interview_korean.wav -m small --translate -l ko

# 7. Multi-Speaker Diarization (Who Spoke When)
termux-stt diarize discussion.wav --speakers 3 --format rttm -o speakers.rttm

# 8. Live Microphone Real-Time Listening (Termux-API / Vosk)
termux-stt listen -e vosk -m small-ko

# 9. Zero-Configuration Built-in Benchmark Demo
termux-stt demo

# 10. Hardware & Driver Diagnostic Doctor
termux-stt doctor

# 11. Model Weight Management
termux-stt models list
termux-stt models download whisper small

3.2 Python SDK

import termux_stt

# 1. Initialize High-Accuracy Whisper Engine with Vulkan GPU Acceleration
engine = termux_stt.create_engine(
    "whisper",
    model="base",
    device="auto",     # "auto", "vulkan", "gpu", or "cpu"
    beam_size=1        # Default: 1 (fast greedy decoding); set 5+ for multi-beam
)

# 2. Transcribe Audio File with Detailed Segment Output
result = engine.transcribe("meeting.wav", lang="en")
print(f"Full Text: {result.text}")
print(f"Duration: {result.audio_duration_sec:.2f}s | Elapsed: {result.elapsed_ms:.1f}ms | RTF: {result.rtf:.4f}x")

for segment in result.segments:
    print(f"[{segment.start_sec:.2f}s -> {segment.end_sec:.2f}s] {segment.text}")

# 3. Voice Activity Detection & Context Prompting
result_vad = engine.transcribe("noisy_lecture.wav", vad=True, prompt="Discussion on quantum computing")
print(f"VAD Result: {result_vad.text}")

# 4. Instant Low-Latency Streaming with Vosk Engine (<30ms Latency)
vosk_engine = termux_stt.create_engine("vosk", model="small-ko")
vosk_result = vosk_engine.transcribe("quick_voice.wav")
print(f"Vosk Output: {vosk_result.text}")

3.3 Node.js / TypeScript SDK

import { createEngine } from 'termux-stt';

async function main() {
  // Initialize Whisper engine with hardware acceleration
  const engine = createEngine('whisper', {
    model: 'base',
    device: 'auto',
    beamSize: 1
  });

  // Transcribe audio file
  const result = await engine.transcribe('meeting.wav', {
    lang: 'en',
    format: 'json'
  });

  console.log('Transcription:', result.text);
  console.log(`Elapsed Time: ${result.elapsedMs}ms | RTF: ${result.rtf}x`);

  // Stream partial results from microphone
  const voskEngine = createEngine('vosk', { model: 'small-ko' });
  voskEngine.on('transcript', (data) => {
    console.log('Live Stream:', data.text);
  });
}

main().catch(console.error);

4. Advanced Architecture & Deep-Dive

Termux-STT features a versatile Tri-Engine architecture designed to adapt dynamically between studio precision and low-latency continuous listening.

flowchart TD
    AudioInput["Audio Input (Microphone / WAV / MP3)"] --> Preproc["Audio Preprocessor (16kHz Mono PCM)"]
    Preproc --> VAD["EnergyVAD / Voice Activity Detector"]
    
    VAD --> Router{"Engine Dispatcher"}
    Router -->|"High Accuracy (GGML Quantized)"| Whisper["WhisperEngine (whisper.cpp Subprocess)"]
    Router -->|"Continuous Streaming (<30ms)"| Vosk["VoskEngine (Kaldi CFFI)"]
    Router -->|"Zipformer Next-Gen ONNX"| Sherpa["SherpaEngine (sherpa-onnx)"]
    
    Whisper & Vosk --> Hybrid["HybridEngine (Speaker Diarization)"]
    Hybrid --> XVec["Vosk 128-d X-Vector Extraction"]
    XVec --> KMeans["Pure-Python K-Means Clustering"]
    
    KMeans --> Export["Export Layer (JSON, SRT, VTT, RTTM)"]

4.1 Subprocess Process Isolation & Mobile Crash Protection

Android Termux environments are prone to out-of-memory kernel kills (OOM) and SIGSEGV segmentation faults during heavy native C++ tensor inference. Termux-STT wraps whisper.cpp and sherpa-onnx in isolated process pools (ProcessPool), intercepting crashes gracefully and returning typed exceptions (AMEVA-STT-E002) without aborting the host Python application.

4.2 Multi-Speaker Diarization Pipeline (No Scikit-Learn Needed)

Traditional speaker diarization requires heavy machine learning frameworks (scikit-learn, torchaudio). Termux-STT integrates an ultra-lightweight HybridEngine:

1. Extracts 128-dimensional acoustic x-vector embeddings using Vosk.
2. Evaluates spatial clustering via an in-house pure-Python K-Means implementation with zero external dependencies.
3. Time-aligns speaker identities with Whisper transcript segments.

import termux_stt

engine = termux_stt.create_engine("hybrid", whisper_model="base", num_speakers=2)
result = engine.transcribe("board_meeting.wav", diarize=True)

for seg in result.segments:
    print(f"[{seg.speaker_id}] {seg.start_sec:.1f}s - {seg.end_sec:.1f}s: {seg.text}")

4.3 Mobile Production Greedy Search (-bs 1) Policy

Autoregressive decoding in Whisper generates text token-by-token. Standard desktop Whisper defaults to Beam Search ($B=5$), which maintains 5 parallel hypothesis states in memory. On mobile Unified Memory Architectures (UMA), this causes severe memory bus saturation, inflating the Key-Value (KV) cache from $49.8 ext{ MB}$ to $249 ext{ MB}$. Termux-STT hardcodes Greedy Search (--beam-size 1 / -bs 1) as the mobile production default, slashing decoder memory bus traffic by 5x while preserving full user sovereignty to override with --beam-size 5.

5. Master Feature & Parameter Matrix

5.1 CLI Subcommands Overview

5.2 Comprehensive Transcription Parameters Matrix

6. Production Code Examples & Diagnostics

6.1 Multi-Agent Voice Pipeline (termux-stt + termux-llamacpp + termux-tts)

Construct a 100% on-device autonomous voice conversational loop:

import termux_stt
import termux_llamacpp as llama
import termux_tts as tts

def run_conversational_cycle(user_audio="input.wav"):
    # 1. Listen & Transcribe User Voice via Termux-STT
    stt = termux_stt.create_engine("whisper", model="base", device="auto", beam_size=1)
    user_text = stt.transcribe(user_audio).text
    print(f"Heard: {user_text}")

    # 2. Reason & Answer via Termux-LlamaCpp
    llm = llama.LlamaRuntime(llama.RuntimeConfig(model_path="qwen2.5-1.5b-instruct"))
    ai_response = llm.generate(prompt=user_text, max_tokens=100)
    print(f"Thought: {ai_response}")

    # 3. Speak Out Loud via Termux-TTS
    with tts.load(engine="vulkan", tier="medium") as voice:
        voice.synthesize(ai_response, output="response.wav")
        print("[SUCCESS] Full on-device voice loop completed.")

if __name__ == "__main__":
    run_conversational_cycle()

6.2 Hardware Diagnostics & Environment Audit

from termux_stt.cli.doctor import run_diagnostics

report = run_diagnostics()
print(f"Vulkan GPU Available: {report.get('vulkan_available')}")
print(f"Optimal Threads: {report.get('optimal_threads')}")
print(f"Installed Engines: {report.get('available_engines')}")

7. Empirical Mobile Hardware Benchmarks (6-SoC Physical Fleet)

7.1 Production Fleet Benchmark Scorecard

Empirical benchmarks conducted on physical Android hardware using JFK's 60.00s 16kHz Mono Inaugural Address (samples/jfk_1min.wav) with Greedy Search (-bs 1) under Vulkan GPU acceleration:

Real-Time Factor (RTF) Definition: $\text{RTF} = \frac{\text{Processing Latency (Seconds)}}{\text{Audio Duration (Seconds)}}$.
An RTF of 0.064x means 60 seconds of recorded speech is transcribed into text in only 3.82 seconds.

7.2 CPU vs. Vulkan GPU Speedup Comparison

Comparative evaluation against ARM NEON 4-thread CPU execution illustrates massive GPU acceleration:

[Throughput Comparison: 60s JFK Audio Processing Speed]
Model: Whisper Small (244M)

Galaxy A35 (Exynos 1380 / Mali-G68 MP5)
  CPU (NEON 4-threads):  |==================================================| 482.1s
  GPU (Vulkan0):         |======| 63.8s  [7.56x Faster]

Galaxy S22 (Snapdragon 8 Gen 1 / Adreno 730)
  CPU (NEON 4-threads):  |========================================| 394.3s
  GPU (Vulkan0):         |=======| 70.9s  [5.56x Faster]

Galaxy S20 (Snapdragon 865 / Adreno 650)
  CPU (NEON 4-threads):  |==================================================| 512.6s
  GPU (Vulkan0):         |============| 128.5s  [3.99x Faster]

Vulkan native GPU execution achieves a $4.0\times\text{ to }7.6\times$ throughput speedup over optimized multi-threaded ARM NEON CPU computation while reducing overall battery drain via the Race-to-Sleep principle.

8. GPU Interconnect Architecture & Compatibility

8.1 Vulkan Compute Acceleration Pipeline

Termux-STT interfaces directly with Android's Bionic Vulkan loader (/system/lib64/libvulkan.so). Tensor mel-spectrogram transformations and encoder attention blocks are offloaded to mobile GPU SPIR-V compute shaders via whisper.cpp Vulkan backend bindings.

8.2 Silicon Compatibility Matrix

• Qualcomm Snapdragon (Adreno 6xx, 7xx, 8xx):
  • Tier-1 Full Support. Native FP16 compute instructions and high dispatch concurrency deliver RTF performance as fast as 0.064x on Snapdragon 8 Elite. Adreno SoftMax workgroups are strictly constrained to hardware subgroup size (64) for rock-solid stability.
• Samsung Exynos / MediaTek Dimensity (ARM Mali / Immortalis):
  • Supported. Mali tile-based architectures benefit from continuous compute queues. Full execution across Tiny, Base, Small, and Turbo without shader stalls.
• Strict Zero-Silent-Fallback:
  • Requesting --device vulkan without valid Vulkan drivers immediately triggers typed exceptions (PlatformNotSupportedError, error code AMEVA-STT-E002), preventing silent fallback to unoptimized CPU execution.

9. Architectural Trade-Off Analysis & Decision Outcomes

In resource-constrained mobile systems engineering, every architectural design represents an explicit compromise between competing physical constraints:

For the comprehensive academic analysis, kernel watchdog expiry mathematical proofs, and driver boundary traces, refer to the Master Technical Treatise (English) and 연구 백서 (Korean).

10. Hardware Requirements & Operational Limits

10.1 Hardware Specifications

10.2 Operational Limits & Best Practices

• 32-Bit ARM (armeabi-v7a): Not supported for Whisper GPU neural inference. Use Vosk CFFI for legacy 32-bit hardware.
• Microphone Permissions: Real-time microphone listening (termux-stt listen) requires Android microphone permission granted to Termux: termux-microphone-record.
• 6GB RAM Mid-Range Devices: For large models (Turbo 809M) on 6GB devices such as Galaxy A53, enable +2GB RAM Plus (zRAM) in Android settings to eliminate page thrashing.

11. 24/7 Unattended Background Execution Guide

Android aggressively kills background user-space processes inside Termux unless battery and process monitor policies are explicitly configured:

11.1 Stage 1: Termux Kernel Wake-Lock

Prevent the mobile CPU from entering low-power sleep states:

termux-wake-lock

11.2 Stage 2: Android GUI Battery Optimization Exemption

1. Open Android Settings > Apps > Termux > Battery.
2. Set battery policy to Unrestricted (Disable power-saving restrictions).
3. Under Permissions, grant Microphone and Display over other apps.

11.3 Stage 3: ADB Phantom Process Killer Exemption (Android 12+)

# Disable Android Phantom Process Killer
adb shell device_config put activity_manager max_phantom_processes 2147483647
adb shell settings put global settings_enable_monitor_phantom_procs false

# Verify configuration
adb shell settings get global settings_enable_monitor_phantom_procs
# Expected output: false

12. Open Source License

Termux-STT is open-sourced under the MIT License.

Copyright (c) 2026 Eunho Kim (@uno-km) & AMEVA Open-Source Foundation.

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

13. SEO Technical Keywords & Ecosystem Metadata

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

Official Documentation & Foundation Ecosystem

• Official Documentation Portal: https://uno-km.github.io/termux-stt/
• GitHub Repository: https://github.com/uno-km/termux-stt
• AMEVA Foundation Portal: https://uno-km.vercel.app/foundation/index.html
• Master Research Treatise (English): docs/research/on_device_vulkan_stt_master_treatise.md
• 연구 백서 (Korean): docs/research/on_device_vulkan_stt_master_treatise_kor.md