naijaVox2_quantized_int8_gguf

naijaVox 2.0

naijaVox 2.0 is a fine-tuned version of openai/whisper-large-v3, optimized for Nigerian English, Pidgin, and other Nigerian languages.

Model Details

Intended Use

• Transcription of Nigerian English, Pidgin English, Yoruba, Igbo, Hausa, and related languages/accents
• General-purpose English ASR with improved performance on West African accents
• Integration into speech-to-text pipelines

How to Use

With Hugging Face Transformers

from transformers import WhisperForConditionalGeneration, WhisperProcessor
import torch

model_id = "amn-raw/naijaVox2-quantized-int8-gguf"

processor = WhisperProcessor.from_pretrained(model_id)
model = WhisperForConditionalGeneration.from_pretrained(
    model_id, torch_dtype=torch.float16, device_map="auto"
)

# Transcribe
from datasets import load_dataset, Audio
ds = load_dataset("your_dataset", split="test")
ds = ds.cast_column("audio", Audio(sampling_rate=16000))

sample = ds[0]["audio"]
input_features = processor(
    sample["array"], sampling_rate=sample["sampling_rate"], return_tensors="pt"
).input_features.to(model.device, dtype=torch.float16)

predicted_ids = model.generate(input_features)
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
print(transcription)

With vLLM (OpenAI-compatible server)

pip install vllm
vllm serve amn-raw/naijaVox2-quantized-int8-gguf \
  --max-model-len 448 \
  --max-num-seqs 100 \
  --task transcription

Then call the transcription endpoint:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
response = client.audio.transcriptions.create(
    model="amn-raw/naijaVox2-quantized-int8-gguf",
    file=open("audio.wav", "rb"),
)
print(response.text)

With transcribe.cpp (GGUF)

Pre-quantized GGUF files are available in this repo under the gguf/ directory.

# Using transcribe.cpp
build/bin/transcribe-cli \
  -m naijavox-2.0-Q8_0.gguf \
  audio.wav

GGUF Quantizations

Training

• Base model: openai/whisper-large-v3
• Fine-tuning framework: Hugging Face Transformers
• Audio format: 16 kHz mono

Limitations

• Optimized for Nigerian languages/accents; may not improve over base whisper-large-v3 on other languages
• 30-second window limit per chunk (long audio is chunked automatically)
• No speaker diarization

License

Apache 2.0 (same as base whisper-large-v3)