Spaces:

---

treble-technologies

/

ffasr

Like 46

Running on CPU Upgrade

Hi Rohan,
please submit the eval script as required so we can run your model. (example given)
Best,
Shivam

Hi Shivam,

Thank you! It looks like the form won’t allow me to resubmit the same model while submission 0f65f283 is pending, so I’ve included the evaluation script below.

For reference, the repository contains the complete fine-tuned checkpoint.

Requirement:

qwen-asr==0.0.6

from pathlib import Path

import soundfile as sf
import torch
from qwen_asr import Qwen3ASRModel

model = Qwen3ASRModel.from_pretrained(
    "rohansheth/tiro-qwen3-asr-1.7b-ffasr-v1",
    dtype=torch.bfloat16,
    device_map="cuda:0",
    max_inference_batch_size=1,
    max_new_tokens=256,
)

def evaluate(file: Path) -> str:
    audio, sr = sf.read(str(file), dtype="float32", always_2d=True)
    audio = audio.mean(axis=1)
    results = model.transcribe(
        audio=(audio, int(sr)),
        language="English",
    )
    return results[0].text

Would you be able to use this with the existing submission, or should it be rejected first so I can resubmit?

Thanks,
Rohan

Hey, sorry I just wanted to follow up on this. Is there anything else you'd need me to do before running the submission?

Hi, your results have been published. :)
Best,
Shivam

Hey, sorry to ping you again! I submitted a revised version of the model recently. I know my last one was a pain to run, so this time I included a tested custom eval script, which should make things much smoother 😅. Thanks!

Model: rohansheth/tiro-qwen3-asr-1.7b-v2

^ @whojavumusic

New results are now published, Congratulations for #1 😄

Thanks so much! Sorry again for all the pings