Downloads · 30 days
52
25% of all-time downloads
Abolfazl27Aflhami88/houshino-whisper-fa
houshino-whisper-fa is a automatic speech recognition model from Abolfazl27Aflhami88. Use it when you need speech turned into text. The card lists the license as apache-2.0.
houshino-whisper-fa is a fine-tuned checkpoint of openai/whisper-small (244M parameters) adapted for informal conversational Persian and regional dialectal variations, including Kurdish and Lori accents.
Downloads · 30 days
52
25% of all-time downloads
All-time downloads
208
Public
Parameters
242M
967 MB on disk
Likes
1
Public
Click a slice to open those files.
.safetensors967 MB · 100%
From the Hugging Face model README
houshino-whisper-fa is a fine-tuned checkpoint of openai/whisper-small (244M parameters) adapted for informal conversational Persian and regional dialectal variations, including Kurdish and Lori accents.
While base Whisper models handle formal, standard written Persian reasonably well, they degrade significantly on informal speech, colloquial contractions, and regional pronunciations. This model was fine-tuned specifically to address that gap.
openai/whisper-smallautomatic-speech-recognition)Before selecting whisper-small, we evaluated several candidate architectures against informal Persian audio samples:
Conversational speech collected from real environments is messy. The initial dataset was roughly 53 GB of raw, uncurated audio.
Feeding this raw data directly into training degraded attention heads due to prolonged silences, low signal-to-noise ratios, clipped audio, and background chatter. We built a dedicated preprocessing and filtering pipeline using Librosa and FFmpeg:
This process filtered the raw 53 GB dump down to a 7.7 GB curated training subset of clear, vocal-dense colloquial Persian, Kurdish, and Lori speech samples.
The model was fine-tuned using PyTorch and Hugging Face transformers. The focus was adjusting the cross-attention and decoder layers so token predictions reflect conversational grammar and colloquial phrasing rather than formal literary Persian.
Evaluated on an informal conversational and dialectal test set:
(For context: off-the-shelf zero-shot base models typically show >50% WER on these informal and dialectal test samples).
import torch
import librosa
from transformers import WhisperProcessor, WhisperForConditionalGeneration
processor = WhisperProcessor.from_pretrained("Abolfazl27Aflhami88/houshino-whisper-fa")
model = WhisperForConditionalGeneration.from_pretrained("Abolfazl27Aflhami88/houshino-whisper-fa")
# Ensure 16kHz mono audio input
audio, sr = librosa.load("path_to_audio.wav", sr=16000)
input_features = processor(audio, sampling_rate=16000, return_tensors="pt").input_features
with torch.no_grad():
predicted_ids = model.generate(input_features, language="persian", task="transcribe")
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
print(transcription)