dicksonsarpong9/qwen3-asr-asante-twi

Qwen3-ASR Asante Twi

Author: Dickson Sarpong

Overview

This model is a fine-tuned Automatic Speech Recognition (ASR) model for Asante Twi, a major Ghanaian language spoken primarily in the Ashanti region of Ghana.

The model is adapted from Qwen3-ASR to improve speech recognition performance on Asante Twi audio, enabling transcription of spoken Twi into text.

Training data

Trained on the Asante Twi subset of the Ghana Speech dataset.

The Ghana Speech dataset contains audio and text data across 42 Ghanaian language subsets and is licensed under CC BY-NC 4.0.

Intended use & license

Non-commercial use only (CC BY-NC 4.0, inherited from the training data).

This model is intended for:

• Research and academic experiments
• Ghanaian language speech technology development
• Automatic speech recognition research
• Low-resource language studies

Commercial usage requires appropriate licensing of the underlying training data.

How to use

import torch
from qwen_asr import Qwen3ASRModel

model = Qwen3ASRModel.from_pretrained(
    "dicksonsarpong9/qwen3-asr-asante-twi",
    dtype=torch.bfloat16,
    device_map="cuda:0",
)

result = model.transcribe(
    audio="sample.wav"
)

print("Language:", result[0].language)
print("Text:", result[0].text)

Training details

• Base model / architecture: Qwen3-ASR
• Language subset: Asante Twi
• Hardware: NVIDIA H200 (Ghana NLP)
• Training epochs: 3
• Training steps: 6,588
• Final training loss: 1.5155
• Precision: bfloat16

Benchmark

Inference test on NVIDIA H200:

Metric	Result
Audio duration	4.48 seconds
Inference latency	1.15 seconds
Real-Time Factor (RTF)	0.256x

The model processed audio approximately 3.9× faster than real-time in this test.

Limitations

• The model is trained specifically on Asante Twi and may not represent other Akan varieties equally.
• Performance may vary with noisy recordings, telephone-quality audio, and heavy code-switching between Twi and English.
• Further evaluation using a dedicated held-out Asante Twi benchmark is recommended.

Acknowledgements

Compute resources provided by AI Skills and Compute Africa (AISCA).

Trained on the Ghana NLP H200 GPU.

Special thanks to the GhanaNLP Community (ghananlpcommunity) for providing the Ghana Speech dataset and supporting African language technology research.

Please keep derivatives non-commercial and share improvements back with the Ghana NLP community.