= 6.9× over open SOTA; trails commercial Deepgram on entity-dense Hindi (0.485) — Hi is the only language in our 3-lang study where commercial has substantial entity coverage.

Read-prose regression vs vasista22 base: +9 pp on FLEURS, +9 pp on CV25, tied on IV.

Usage

python

from transformers import WhisperForConditionalGeneration, WhisperProcessor
from peft import PeftModel
import torch, librosa

base_model = "vasista22/whisper-hindi-large-v2"
processor = WhisperProcessor.from_pretrained(base_model, language="hindi", task="transcribe")
model = WhisperForConditionalGeneration.from_pretrained(base_model, torch_dtype=torch.bfloat16).to("cuda")
forced = processor.tokenizer.get_decoder_prompt_ids(language="hindi", task="transcribe")
model.config.forced_decoder_ids = forced
model.generation_config.forced_decoder_ids = forced
model.generation_config.suppress_tokens = []

model = PeftModel.from_pretrained(model, "Praxel/praxy-stt-hi-rb")
model.eval()

audio, _ = librosa.load("path/to/audio.wav", sr=16000, mono=True)
feats = processor.feature_extractor(audio, sampling_rate=16000, return_tensors="pt").input_features.to("cuda", dtype=torch.bfloat16)
pred_ids = model.generate(feats, max_new_tokens=400, num_beams=1)
print(processor.tokenizer.decode(pred_ids[0], skip_special_tokens=True).strip())

Training

LoRA rank 16, alpha 32, on q/k/v/out_proj of vasista22/whisper-hindi-large-v2. 4000 steps Modal A10G. Cartesia rows held out for evaluation. Pin chain: transformers==4.36.2, peft==0.10.0.

Companion work + paper

Praxy Voice TTS (paper #1): arXiv:2604.25441
PSP (paper #2): arXiv:2604.25476
STT Flywheel (this paper): arXiv:2605.03073; code at github.com/praxelhq/stt-flywheel

License: Apache-2.0.