Brahvi-Whisper-Small

Brahvi Whisper Small

Brahvi Whisper Small is a full-parameter fine-tune of openai/whisper-small for automatic speech recognition in Pakistani Brahvi (also known as Brahui). It produces Perso-Arabic Brahvi transcriptions from real-world speech.

To the best of the author's knowledge as of September 16, 2026, this is the world's first publicly released automatic speech recognition model dedicated to the Brahvi language. This claim concerns dedicated, publicly documented model releases and does not exclude unpublished systems or broad multilingual models that may provide incidental Brahvi coverage.

The model was trained with TBOGamer22/BrahuiSpeech-70H-V2.

Important Prompt Setting

Whisper does not provide a dedicated brh language prompt token. This model was therefore trained with Whisper's Urdu token while the target text remained Brahvi. Use language="ur" and task="transcribe". The model was not trained for speech translation.

Usage

import math

import numpy as np
import soundfile as sf
import torch
from scipy.signal import resample_poly
from transformers import WhisperForConditionalGeneration, WhisperProcessor

model_id = "TBOGamer22/Brahvi-Whisper-Small"
use_cuda = torch.cuda.is_available()
dtype = torch.float16 if use_cuda else torch.float32

processor = WhisperProcessor.from_pretrained(model_id)
model = WhisperForConditionalGeneration.from_pretrained(
    model_id,
    dtype=dtype,
    low_cpu_mem_usage=True,
    use_safetensors=True,
).to("cuda" if use_cuda else "cpu")

audio, sample_rate = sf.read("brahvi_audio.wav", dtype="float32", always_2d=True)
audio = audio.mean(axis=1)
if sample_rate != 16000:
    divisor = math.gcd(sample_rate, 16000)
    audio = resample_poly(
        audio,
        16000 // divisor,
        sample_rate // divisor,
    ).astype(np.float32, copy=False)

inputs = processor(
    audio,
    sampling_rate=16000,
    return_tensors="pt",
)
input_features = inputs.input_features.to(model.device, dtype=dtype)

with torch.inference_mode():
    predicted_ids = model.generate(
        input_features,
        language="ur",
        task="transcribe",
    )

text = processor.batch_decode(
    predicted_ids,
    skip_special_tokens=True,
    clean_up_tokenization_spaces=False,
)[0]
print(text)

Training

• Base model: openai/whisper-small
• Training method: full-model fine-tuning
• Dataset: TBOGamer22/BrahuiSpeech-70H-V2
• Training examples: 14,978 clips, 66 hours 39 minutes 29 seconds
• Held-out evaluation: 648 clips, 2 hours 52 minutes 43 seconds
• Split seed: 42
• Epochs: 5
• Final checkpoint: step 9,365
• Learning rate: 1e-5
• Scheduler: cosine with 500 warmup steps
• Per-device batch size: 8
• Precision: BF16 on an NVIDIA RTX 3090
• Objective: Brahvi transcription only
• Prompt: Whisper Urdu language token (ur) with task transcribe
• Augmentation: telephony/GSM-style augmentation, RIRS point-source noise, and 8 kHz bandwidth simulation

Evaluation

Results were measured on the fixed 648-clip held-out subset. WER and CER are case- and punctuation-sensitive after whitespace normalization. This is an in-domain held-out evaluation, not an independent external benchmark.

The complete per-epoch results and the released weight hash are provided in training_results.json.

Intended Use

This model is intended for Brahvi ASR research, transcription experiments, low-resource speech technology, and continued fine-tuning. Predictions should be reviewed by a fluent Brahvi speaker before publication or use in consequential workflows.

Limitations

• Performance varies with dialect, recording quality, background noise, speaker, code-switching, and domain.
• The evaluation subset comes from the same dataset as the training split and may not represent unseen recording sources.
• The Urdu prompt token is a compatibility choice and does not mean the model is intended to transcribe Urdu.
• The model can omit, repeat, or hallucinate words, especially in silence or severely degraded audio.
• The training corpus contains diverse real-world media, so users should review the dataset card and applicable usage rights before deployment.

Release Integrity

The released weights are the epoch-5 final export and are byte-identical to checkpoint-9365/model.safetensors from training.