🎤 General-Purpose Swahili ASR Model (Open Datasets + Domain Data)
This automatic speech recognition (ASR) model is trained using open multilingual datasets and a multi-domain in-house dataset to provide high-accuracy transcription for clean, read-aloud Swahili speech.
It achieves strong generalization, maintaining benchmark accuracy while improving performance on real-world test data. The model can do well for both clean and noisy audios.
This model is part of a full ASR ablation study that analyzes and understands the robustness of data and in dealing with different modes and variations of data collections. 👉 View all models on GitHub
We are particularly interested in validating the conclusions we’ve observed through our ablation studies:
While benchmark datasets like FLEURS are useful for comparison, they do not fully capture the variability and challenges of real-world speech — especially for underrepresented languages like Swahili. We are inviting the community to try out these models and help assess:
- How well the models perform on natural, conversational, or noisy audio
- Open-source datasets (like Common Voice & FLEURS) perform well on clean, benchmark speech.
- Whether the improvements we've seen in combining diverse datasets generalize to your use case
- Gaps between benchmark results and real-world usability
- A combination of both yields balanced results but depends on data quality and label accuracy.
Model
Whisper is a general-purpose speech recognition model. It is trained on a large dataset of diverse audio and is also a multitasking model that can perform multilingual speech recognition, speech translation, and language identification.
🚀 How to Use
from transformers import WhisperForConditionalGeneration, WhisperProcessor
from transformers import pipeline
from transformers.utils import is_flash_attn_2_available
import torch
processor = WhisperProcessor.from_pretrained("openai/whisper-large-v2")
model = WhisperForConditionalGeneration.from_pretrained("RafatK/Swahili-Whisper_Largev2-Decodis_Comb_FT", torch_dtype=torch.float16).to("cuda")
# model.generation_config.input_ids = model.generation_config.forced_decoder_ids
model.generation_config.forced_decoder_ids = None
forced_decoder_ids = processor.get_decoder_prompt_ids(language="swahili", task="transcribe")
pipe = pipeline(
"automatic-speech-recognition",
model=model,
tokenizer=processor.tokenizer, # Explicitly pass the tokenizer
feature_extractor=processor.feature_extractor, # Explicitly pass the feature extractor
chunk_length_s=15,
device=0,
model_kwargs={"attn_implementation": "flash_attention_2"} if is_flash_attn_2_available() else {"attn_implementation": "sdpa"},
generate_kwargs = {
'num_beams':5,
'max_new_tokens':440,
'early_stopping':True,
'repetition_penalty': 1.8,
'language': 'swahili',
'task': 'transcribe'
}
)
text_output = pipe("audio.wav")['text']
📊 Total Duration: ~400 hours
📁 Languages: Swahili (sw)
🏋️♂️ Training Strategy
- Architecture:
whisper-large-v2 - Framework: Whisper and Huggingface Transformers
- Sampling rate: 16 kHz
- Preprocessing: Volume normalization, High-Grade noise addition, Prosodic Augmentation, silence trimming
- Learning Rate: 1e-5
- Optimizer: Adamw_pytorch
- Steps: 3000
- Pretrained on open data
- Fine-tuned on domain data
📈 Evaluation Metric (WER)
| Dataset | This Model | Whisper Large V2 |
|---|---|---|
| FLEURS (benchmark) | 12.41 | 39.40 |
| Decodis Test Set (Collected by DECODIS) | 39.42 | 99.98 |
🎯 Intended Use
- General-purpose transcription systems
- Balanced performance on clean and noisy data
- Speech interfaces in multilingual and informal settings
⚠️ Limitations
- Slight trade-off in benchmark precision
- May need more domain data for extreme acoustic variation
📝 Please try the models and share your feedback, issues, or results via:
GitHub Issues: Submit an issue
Hugging Face Discussions: Join the conversation
Your feedback will help us refine our dataset and improve ASR for underrepresented languages like Swahili and Yoruba.