Advanced Parameters
Detailed handbook for fine-tuning performance, latency, and hardware utilization.
EngineConfig Parameters
| Parameter | Type | Default | Description |
|---|---|---|---|
engine |
str | "whisper" |
STT engine backend: "whisper", "vosk", "sherpa", "hybrid". |
model |
str | "base" |
Model size or identifier (e.g. "tiny", "base", "small", "medium"). |
lang |
str | "ko" |
ISO 639-1 language code ("ko", "en", "ja", "zh", "auto"). |
threads |
int | None (Auto) |
Number of CPU threads. Auto-detects Big-core count (e.g. 4 for Exynos 1380). |
vad |
bool | True |
Enable Voice Activity Detection for silence filtering and chunking. |
vad_threshold |
float | 0.5 |
VAD sensitivity threshold between 0.0 (aggressive) and 1.0 (conservative). |
quantization |
str | "q5_1" |
GGML quantization level: "f16", "q8_0", "q5_1", "q4_0". |
num_speakers |
int | 0 |
Number of speakers for diarization. 0 disables diarization; 2+ enables clustering. |
Hardware Pinning & Big-Cores
Modern mobile SoCs (Exynos, Snapdragon, Dimensity) use big.LITTLE architectures. termux-stt automatically binds inference to high-performance Big cores (e.g., Cortex-A78) for maximum RTF.
from termux_stt.platform.hardware import detect_hardware, get_optimal_threads
info = detect_hardware()
print("CPU:", info.cpu_model)
print("Big Cores:", info.big_cores)
print("NEON Support:", info.neon_support)
print("Optimal Threads:", get_optimal_threads())