Pooya-Fallah/whisper-medium-fa

Persian Whisper Medium

A Persian speech-recognition checkpoint based on openai/whisper-medium. This repository contains the inference weights, model configuration, and saved processor/tokenizer. It does not contain training audio or optimizer state.

Access

The model card is public and model downloads require manual approval. Sign in on this model page, request access, and wait for the owner to approve your account before downloading.

Authenticate on the machine where you will run inference using your own Hugging Face account and a token with read access to this repository:

hf auth login

Installation and inference

Use Python 3.10 or newer and a PyTorch build suitable for your CPU or CUDA GPU:

python -m pip install 'transformers>=5.9,<6' torch librosa soundfile huggingface_hub

Save the following as transcribe.py, put an audio file at speech.wav, then run python transcribe.py. The example reads the token saved by hf auth login. For reproducibility, replace main with a commit SHA from this repository.

import librosa
import torch
from transformers import WhisperForConditionalGeneration, WhisperProcessor

repo_id = "Pooya-Fallah/whisper-medium-fa"
revision = "main"
device = "cuda" if torch.cuda.is_available() else "cpu"
processor = WhisperProcessor.from_pretrained(repo_id, revision=revision, token=True)
model = WhisperForConditionalGeneration.from_pretrained(
    repo_id, revision=revision, token=True
).to(device).eval()

# librosa converts to mono and resamples to the required 16 kHz.
audio, _ = librosa.load("speech.wav", sr=16000, mono=True)
if len(audio) == 0:
    raise ValueError("Audio is empty")

# This simple example handles one recording of at most 30 seconds.
if len(audio) > 30 * 16000:
    raise ValueError("Use the long-audio example below for recordings over 30 seconds")
inputs = processor(
    audio, sampling_rate=16000, return_tensors="pt", return_attention_mask=True
).to(device)
with torch.inference_mode():
    ids = model.generate(**inputs, language="fa", task="transcribe")
print(processor.batch_decode(ids, skip_special_tokens=True)[0])

For recordings longer than 30 seconds, replace the inputs = ... block and the length guard above with Whisper's sequential long-form generation:

inputs = processor(
    audio, sampling_rate=16000, return_tensors="pt",
    return_attention_mask=True, truncation=False, padding="longest"
).to(device)
with torch.inference_mode():
    ids = model.generate(
        **inputs, language="fa", task="transcribe", return_timestamps=True
    )
print(processor.batch_decode(ids, skip_special_tokens=True)[0])

Long recordings require more memory. Run inference on your own machine or GPU server; storing weights on the Hub does not itself provision an inference API.

Training provenance

This export is from the whisper-medium-fa run. Its saved configuration lists Common Voice, degraded and long variants, FarsSpon, and FLEURS. Coverage of "all data" has not been verified against the actual training manifests and logs.

Selected export: final.

Datasets listed in the saved configuration:

• cv-corpus-25.0
• cv-corpus-25.0-degraded-v2
• FarsSpon_train_dev
• cv-corpus-25.0-long
• FarsSpon_train_dev-long
• cv-corpus-25.0-degraded-long-v2
• fleurs-normalized

These values describe the saved configuration, not an independently audited record of every training sample. publication.json records the model key and export name, selected inference files, and the saved training-configuration SHA-256.

Evaluation and limitations

No WER/CER scores have been verified for this exact published export. No claim is made that this checkpoint outperforms another run or the base model. Validate it on your intended domain before use. Noise, accents, overlapping speakers, music, and silence can produce errors or hallucinated text. Review transcriptions before using them as research annotations.

License

Consult the base model for its license. A redistribution license for this fine-tuned checkpoint has not been declared here; training-data rights and the intended license should be reviewed by the model owner before adding license metadata.