OzLabs

/

Caspi-1.7B

like 4

Follow

Oz Labs 4

this is my code 😀

import torch
from qwen_asr import Qwen3ASRModel

configure_cpu_threads(threads)
device = "cpu"
print("[1/2] Loading Caspi-1.7B...", flush=True)
load_started = time.perf_counter()
try:
    model = Qwen3ASRModel.from_pretrained(
        CASPI_MODEL,
        dtype=torch.float32,
        device_map=device,
        max_inference_batch_size=1,
    )
except Exception as error:
    raise ModelLoadError(str(error)) from error
load_time = time.perf_counter() - load_started

print("[1/2] Transcribing with Caspi...", flush=True)
transcription_started = time.perf_counter()
try:
    transcription = model.transcribe(audio=str(wav_path), language="Hebrew")
    transcript = transcription[0].text
finally:
    del model

Drop the language= argument:

transcription = model.transcribe(audio=str(wav_path))
transcript = transcription[0].text

Since Caspi is fine-tuned exclusively on Hebrew, the forced-language prefix has nothing to add to the model outputs Hebrew regardless.