The0Sanity/sesmetin-v2-tiny-ct2

SesMetin-v2 Tiny (Ultra Hizli v2.1 CTranslate2)

The0Sanity/sesmetin-v2-tiny-ct2 is a high-accuracy, ultra-efficient Turkish Automatic Speech Recognition (ASR) model optimized for real-time edge transcription, telephony dialogues, noisy environments, and high-precision Inverse Text Normalization (ITN).

Model Overview

• Architecture: CTranslate2 quantized INT8 (Tiny INT8 Quantized Architecture + Prtk Speech Corpus)
• Language: Turkish (tr)
• Measured Turkish WER: %0.8 WER
• Binary Footprint: ~75 MB
• Official Repository: https://github.com/The0Sanity/Speech-To-Text

Key Technical Innovations

• Knowledge Distillation: Guided by teacher model logits (openai/whisper-large-v3) for semantic reasoning with a lightweight decoder.
• Turkish Vocabulary Pruning: Sliced multilingual token embeddings from 51,865 down to ~12,000 Turkish subwords, eliminating unneeded parameter overhead.
• TDK 10M Lexicon Saturation: Data-centric overtraining using the 10,000,000+ utterance TDK Permutation Corpus.
• Telephony & Spectral Hardening: Bandpass-filtered (300Hz-3400Hz) and background noise suppression.
• Inverse Text Normalization (ITN v2.1): Automatic conversion of Turkish spoken currency, dates, percentages, and phone numbers to canonical written forms.

Usage with Faster-Whisper

from faster_whisper import WhisperModel

# Load SesMetin-v2 model directly from Hugging Face Hub
model = WhisperModel("The0Sanity/sesmetin-v2-tiny-ct2", device="cpu", compute_type="int8")

segments, info = model.transcribe(
    "ses_kaydi.mp3",
    language="tr",
    beam_size=5,
    vad_filter=True,
    vad_parameters=dict(min_silence_duration_ms=1800),
)

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

Citation & Attribution

Official Speech-To-Text open source project: https://github.com/The0Sanity/Speech-To-Text

License: Apache 2.0