Kennethdot/kasanoma_wav2vec2

English–Twi Code-Switching ASR Model — Kasanoma (wav2vec2)

Model Overview

This repository contains a fine-tuned checkpoint of facebook/wav2vec2-large-xlsr-53 for English–Twi code-switching speech transcription. It is further fine-tuned on a realistic bilingual dataset containing English & Twi mixed-language utterances.

The model supports natural bilingual speech, including intra-sentential and inter-sentential code-switching.

How to Use

import torch
from datasets import load_dataset, Audio
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

model = Wav2Vec2ForCTC.from_pretrained("Kennethdot/kasanoma_wav2vec2")
processor = Wav2Vec2Processor.from_pretrained("Kennethdot/kasanoma_wav2vec2")

device = "cuda" if torch.cuda.is_available() else "cpu"
model = model.to(device)
model.eval()

# Load a sample from the test set
dataset = load_dataset(
    "Kennethdot/Ghana_English-Twi_Code-switching_ASR",
    split="test"
).cast_column("audio", Audio(sampling_rate=16000))

sample = dataset[0]["audio"]

inputs = processor(
    sample["array"],
    sampling_rate=sample["sampling_rate"],
    return_tensors="pt",
    padding=True
).to(device)

with torch.no_grad():
    logits = model(**inputs).logits

predicted_ids = torch.argmax(logits, dim=-1)

transcription = processor.batch_decode(
        predicted_ids,
        group_tokens=True,
        skip_special_tokens=False
        )[0].strip()

print(transcription)

Model Details

Property	Value
Task	English–Twi code-switching speech transcription
Base model	facebook/wav2vec2-large-xlsr-53
Fine-tuning dataset	Kennethdot/Ghana_English-Twi_Code-switching_ASR
Dataset size	~100 hours of English–Twi code-switched speech
Sampling rate	16,000 Hz

The dataset was normalised to remove punctuations (, ? . ! ; : " % ") that may destabilize training.

Evaluation Results

Model	CS WER	Twi WER	English WER
Zero-shot XLSR-53	90.39	85.08	110.26
Fine-tuned Model (Kasanoma)	6.58	99.44	100.43

Note: The high monolingual WER scores reflect that this model is optimised for code-switched input. For purely Twi or purely English audio, a monolingual model is likely more appropriate.

Examples

The model produces fluent bilingual outputs with natural speech patterns:

• Example 1 — Ma yɛnkɔgye yɛn ani, it has been a long week.
• Example 2 — Adwuma no yɛ den dodo, I need a vacation.
• Example 3 — Nsuomnam yɛ dɛ paa, w'atry-i grilled tilapia?

Limitations

The model performs well on English–Twi mixed speech. Keep the following in mind:

• Input length: wav2vec2 processes raw waveforms directly but memory usage scales with audio length. For long recordings, apply sliding-window chunking.
• Out-of-distribution input: Performance may degrade on slang, idioms, informal Twi, spelling variation, proper names, or utterances far outside the training distribution.
• Monolingual speech: The model is not optimised for purely English or purely Twi utterances.

Human review is recommended for high-stakes use cases.

Ethical Considerations

• Intended for research and educational use only
• Should not be used for surveillance or unauthorized speech monitoring
• Bias may exist due to dataset imbalance between languages