GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Mannon
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Mannon
Model Tree
openai/whisper-medium
Input Modalities
Output Modalities
Supported Functionality
bash
pip install -U transformers accelerate librosapython
from transformers import pipeline asr = pipeline( "automatic-speech-recognition", model="Mannon/uzbek_stt_v1", chunk_length_s=30, # uzun audiolarni bo‘laklarga bo‘lib ishlaydi device="cuda", # GPU bo‘lmasa: device="cpu") out = asr("audio.wav", generate_kwargs={"language": "uz", "task": "transcribe"})print(out["text"])python
import librosafrom transformers import WhisperProcessor, WhisperForConditionalGeneration repo = "Mannon/uzbek_stt_v1"processor = WhisperProcessor.from_pretrained(repo)model = WhisperForConditionalGeneration.from_pretrained(repo) speech, _ = librosa.load("audio.wav", sr=16_000)features = processor(speech, sampling_rate=16_000, return_tensors="pt").input_features ids = model.generate(features, language="uz", task="transcribe")print(processor.batch_decode(ids, skip_special_tokens=True)[0])| Xususiyat | Qiymat |
|---|---|
| Asosiy model | openai/whisper-medium (~769M parametr) |
| Til | O‘zbek (uz) |
| Chiqish yozuvi | Lotin |
| O‘qitish ma’lumoti | ≈ 1 600 soat o‘zbekcha audio |
| Aniqlik formati | BF16 (o‘qitishda) |
| Audio talabi | 16 kHz, mono |
| Litsenziya | Apache 2.0 |
Rus tilidan o‘zlashgan so‘zlar lotin yozuvida beriladi: brat, davay, prosto va h.k.
8 ta turli test to‘plamidagi 1 864 ta namuna bo‘yicha:
text
Umumiy WER ██████████████████░░░░░░░░░░░░ 16.7 %Umumiy CER ████████░░░░░░░░░░░░░░░░░░░░░░ 7.0 %| Audio turi | WER |
|---|---|
| Toza nutq | ~6–11 % |
| Shovqinli / augmentatsiyalangan | ~12–24 % |
| Shevalar | ~16–25 % |
WER — so‘zlar xatolik darajasi, CER — belgilar xatolik darajasi (qanchalik past bo‘lsa, shunchalik yaxshi).
Model bosqichma-bosqich (curriculum learning) yondashuvi bilan uch bosqichda o‘qitilgan:
text
1. Asos ──▶ 725 soat — tilning asosiy akustik va leksik bazasi2. Barqarorlik ──▶ 394 soat — shovqin va turli yozuv sharoitlariga chidamlilik3. Domenga moslash ──▶ 474 soat — real hayotiy nutq turlariga moslashtirishfrom transformers import pipeline asr = pipeline( "automatic-speech-recognition", model="Mannon/uzbek_stt_v1", chunk_length_s=30, # uzun audiolarni bo‘laklarga bo‘lib ishlaydi device="cuda", # GPU bo‘lmasa: device="cpu") out = asr("audio.wav", generate_kwargs={"language": "uz", "task": "transcribe"})print(out["text"])import librosafrom transformers import WhisperProcessor, WhisperForConditionalGeneration repo = "Mannon/uzbek_stt_v1"processor = WhisperProcessor.from_pretrained(repo)model = WhisperForConditionalGeneration.from_pretrained(repo) speech, _ = librosa.load("audio.wav", sr=16_000)features = processor(speech, sampling_rate=16_000, return_tensors="pt").input_features ids = model.generate(features, language="uz", task="transcribe")print(processor.batch_decode(ids, skip_special_tokens=True)[0])Umumiy WER ██████████████████░░░░░░░░░░░░ 16.7 %Umumiy CER ████████░░░░░░░░░░░░░░░░░░░░░░ 7.0 %1. Asos ──▶ 725 soat — tilning asosiy akustik va leksik bazasi2. Barqarorlik ──▶ 394 soat — shovqin va turli yozuv sharoitlariga chidamlilik3. Domenga moslash ──▶ 474 soat — real hayotiy nutq turlariga moslashtirish