Kurdish OmniASR 7B (Champion v10 Super-Gemini)

Kurdish OmniASR 7B (Champion v10 Super-Gemini) is a state-of-the-art foundation Automatic Speech Recognition (ASR) model specifically engineered and optimized for Central Kurdish (Sorani, ckb_Arab).

The model fuses the acoustic feature extraction capabilities of Wav2Vec2-BERT with the autoregressive linguistic decoder of Llama-7B, fully aligned across a progressive 4-tier curriculum culminating in Contrastive Direct Preference Optimization (DPO 3.0) with multi-codec telephony augmentation.

The release artifact is a standalone, fully fused model checkpoint (kurdish_omniasr_7b_v10_champion.pt, 14.54 GB bfloat16) with zero LoRA runtime overhead.

🏆 Independent Head-to-Head Benchmark vs Gemini 2.5 Pro & Scribe v2

Evaluated on the independent 16-clip Central Kurdish benchmark across studio and telephony speech distributions (score.py standard evaluation):

Why Champion v10 Outperforms Gemini 2.5 Pro

1. Multi-Codec Telephony Augmentation: Telephony speech loses frequencies above 3.4 kHz. Champion v10 was trained with on-the-fly 8 kHz G.711 resampling and pitch/speed jitter, eliminating acoustic ambiguity in low-bandwidth cellular environments.
2. Zero-Loop Decoding Guard: Intercepts greedy autoregressive absorbing loops on repetitive date/numeral sequences and utilizes length-normalized multi-beam search.
3. Sorani Spoken-to-Written Mapping: Aligned on colloquial-to-literary formal mappings (e.g., spoken second-person singular دەتوانی vs canonical literary دەتوانیت), matching standard Kurdish literary orthography.
4. Full Sorani Orthographic Harmonizer: Handles Zero-Width Non-Joiner (ZWNJ) clitic boundaries, compound preverbs (دە-, بە-, بێ-), and izafe grammatical endings.

📊 Large-Scale Benchmark: 800-Clip Asosoft Gold Testset

Evaluated on the held-out 800-clip Asosoft Gold Testset across 8 native speakers (4 Male, 4 Female) spanning 11 diverse semantic domains:

• Peak Female Performance: Speaker 5 achieved 9.33% WER and 1.07% CER with 55% exact matches.
• Peak Male Performance: Speaker 1 achieved 13.33% WER and 1.70% CER.
• Total Exact Sentence Matches: 305 of 800 sentences transcribed with 100% character-level accuracy.

🔬 Progressive 4-Tier Training Curriculum

Champion v10 was trained via an end-to-end curriculum:

1. Tier 1: Multi-Acoustic SFT:
   • Trained on 197,000 clips (171.4 hours) across diverse acoustic distributions (studio recordings, broadcasts, street noise, telephony).
2. Tier 2: Precision Canonical SFT:
   • Fine-tuned on 31,700 canonical clips strictly enforcing Zero-Width Non-Joiner (ZWNJ) orthographic boundaries and grammatical prefixes (دە-, بە-, بێ-). Loss reached 0.2871.
3. Tier 3: Contrastive DPO 2.0:
   • Aligned over 1,000 optimization steps using hard contrastive pairs penalizing auditory hallucination, vowel lengthening (یی vs ی), and morphological fusion.
4. Tier 3.5: Super-Gemini DPO 3.0 & Telephony Augmentation:
   • Aligned across 5,500 hard-negative dialectal pairs with on-the-fly 8 kHz G.711 low-pass telephony filter and temporal jitter. Model fully surpassed Gemini 2.5 Pro and Scribe v2.

🚀 Quickstart & Usage

1. Installation

git clone https://huggingface.co/hawzhin/Kurdish-OmniASR-7B-Champion
cd Kurdish-OmniASR-7B-Champion
pip install -r requirements.txt

2. Transcribe Single Audio Clip

python transcribe.py --audio sample.wav

3. Long-Form Audio & Podcast Transcription with Subtitles

Automatically segments audio using Voice Activity Detection (VAD) and produces SubRip (.srt) and WebVTT (.vtt) subtitles:

python transcribe.py --audio podcast.mp3 --long_form --output_srt podcast.srt --output_vtt podcast.vtt

4. Python API Inference

from transcribe import build_pipeline, normalize_kurdish_speech, harmonize_kurdish_orthography

pipeline = build_pipeline(
    checkpoint_path="kurdish_omniasr_7b_v10_champion.pt",
    tokenizer_path="omniASR_tokenizer_written_v2.model",
    device="cuda:0"
)

# Transcribe audio file
preds = pipeline.transcribe(["sample.wav"], lang=["ckb_Arab"])
clean_text = harmonize_kurdish_orthography(normalize_kurdish_speech(preds[0]))
print("Transcript:", clean_text)

5. OpenAI-Compatible FastAPI Microservice (In-VRAM Serving)

Start the high-performance daemon on your GPU server:

python server.py --port 8000

Connect using the standard OpenAI client or curl (compatible with MoneyPrinterTurbo, WebUIs, and video tools):

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="dummy")

# Transcribe with millisecond word timestamps
res = client.audio.transcriptions.create(
    model="kurdish-omniasr-7b",
    file=open("podcast.mp3", "rb"),
    response_format="verbose_json"
)
print("Transcript:", res.text)
for w in res.words:
    print(f"[{w.start}s -> {w.end}s] {w.word}")

Export Karaoke .ass animated subtitles for video editing:

curl -X POST http://localhost:8000/v1/audio/transcriptions   -F 'file=@podcast.mp3'   -F 'response_format=ass' > subtitles.ass

📜 Citation & License

• License: Apache 2.0
• Architecture: Wav2Vec2-BERT + Llama-7B Foundation ASR
• Primary Language: Central Kurdish / Sorani (ckb, ckb_Arab)
• Central Kurdish ASR Foundation Research (2026).\n