mihu-community-stt-tr

Real-time, low-latency speech-to-text (STT) for Turkish — on the CPU.

Turkish-only · streaming · CPU-native · runs on the edge. A compact 66 M-parameter causal streaming model, built for voice agents, call centers, and on-device transcription where a GPU is not an option — no API key, no cloud, no GPU.

• 🇹🇷 Turkish-first — every parameter serves one language, not ~100
• ⚡ Streaming — partial results as you speak (~320 ms), no 30-second windows
• 🖥️ CPU-native — RTF 0.128 on a single CPU thread (~8× faster than real time)
• 🔌 Tiny — ~68 MB; runs on a Raspberry-Pi-class device
• 🏷️ Entity/brand aware — tuned for Turkish brands, models, and numbers

Installation

pip install mihu-community-stt-tr

Usage

import mihu_stt

# one-shot: transcribe a file (16 kHz mono, or anything soundfile reads)
text = mihu_stt.transcribe("audio.wav")
print(text)

Streaming — feed audio as it arrives and read an updating transcript, with turn-end detection built in:

import mihu_stt

stt = mihu_stt.StreamingSTT()
# pcm16 = 16-bit PCM bytes from your mic / phone line
text, is_final = stt.accept_pcm16(pcm16, sample_rate=16000)
# `text` updates while speech continues; `is_final` turns True at end-of-utterance

Model weights download automatically from Hugging Face (mihuai/mihu-community-stt-tr) the first time and are cached.

Integrations

Drop-in adapters for the common voice-agent stacks (see examples/):

LiveKit Agents — examples/livekit_stt.py:

from livekit.agents import AgentSession
from livekit_stt import MihuSTT
session = AgentSession(stt=MihuSTT(), llm=..., tts=...)

Pipecat — examples/pipecat_stt.py:

from pipecat_stt import MihuSTTService
pipeline = Pipeline([transport.input(), vad, MihuSTTService(), llm, tts, transport.output()])

Benchmarks

Measured in-house under one matched Turkish normalization. Real-human held-out FLEURS-TR (743 clips) and a second real-human set MediaSpeech-TR (200-clip subset), both audited absent from training. Latency on a single CPU thread (AMD EPYC 7282).

Among streaming, CPU-deployable systems mihu-community-stt-tr more than halves Vosk's error on both real-human sets, matches the four-times-larger offline Whisper-small, and is the only option that is real-time on CPU — it also degrades more gracefully than Vosk under additive noise. A detailed technical report — full methodology, ablations, latency/concurrency study, and the leakage audits — is being prepared and will be shared.

Reproduce the numbers

The evaluation set, references, and scripts are in eval/ — including the text (n-gram) and acoustic (chromaprint) contamination audits used to verify the held-out sets are absent from training.

pip install "mihu-community-stt-tr[eval]"
python eval/run_eval.py --data-dir eval/tts-synthetic

License

mihu Community License — free for research, personal, and limited commercial use. See LICENSE for the full terms. Weights are released under the same license on Hugging Face.