Modelpublic

Congo-digital-service/whisper-small-lingala-qlora-merged

sourceHugging Faceapache-2.0updated 18d agoView on Hugging Face

0likes118downloads

Model Card

🇨🇬 whisper-small-lingala-qlora-merged

An openai/whisper-small model adapted to Lingala through QLoRA/LoRA fine-tuning, developed by Congo Digital Services (CDS). This model converts spoken Lingala audio into text (automatic transcription).

🔗 Sister model: LoRA adapters only are available at `Congo-digital-service/whisper-small-lingala-qlora-adapters`

Model Details

Model Description

—Developed by: Congo Digital Services (CDS SARL) — congo-digital.com
—Base model: openai/whisper-small
—Adaptation method: QLoRA/LoRA fine-tuning
—Language: Lingala (ln)
—Model type: Automatic Speech Recognition (ASR), merged (adapters fused into base weights)
—Functional role: Converts Lingala speech recordings into text
—Evaluation metrics: Word Error Rate (WER), Character Error Rate (CER)

Model Sources

—Organization: Congo-digital-service on Hugging Face
—Contact: info@congo-digital.com

Uses

Direct Use

This model can be used directly to transcribe Lingala speech audio into text, either through the transformers pipeline or via a compatible inference server exposing an OpenAI-style transcription endpoint (see Deployment Specification below).

Out-of-Scope Use

This model is not validated for production deployment yet (see Validation Status below), and should not be relied on for high-stakes transcription (legal, medical, or safety-critical audio) without human review.

How to Get Started with the Model

python

from transformers import WhisperForConditionalGeneration, WhisperProcessor
import torch, torchaudio

model_id = "Congo-digital-service/whisper-small-lingala-qlora-merged"

processor = WhisperProcessor.from_pretrained(model_id)
model = WhisperForConditionalGeneration.from_pretrained(model_id)

waveform, sr = torchaudio.load("audio.wav")
inputs = processor(waveform.squeeze().numpy(), sampling_rate=sr, return_tensors="pt")

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

Training Details

Training Procedure

—Training steps: 2,500 steps, evaluated every 250 steps
—Checkpoint selection: best checkpoint selected based on WER, followed by adapter merging and final merged model export

Training Results

Validation loss decreased steadily from 1.3415 to 0.8975, while WER and CER improved progressively through step 2,500 with no clear sign of overfitting:

Step	Training Loss	Validation Loss	WER	CER
250	1.2962	1.3415	64.45%	26.53%
500	1.2138	1.1822	60.29%	24.74%
750	1.1461	1.0893	57.63%	23.60%
1000	1.0547	1.0311	57.21%	23.80%
1250	0.9589	0.9904	56.12%	23.36%
1500	0.9127	0.9625	54.63%	22.40%
1750	0.8317	0.9360	54.45%	22.31%
2000	0.8505	0.9136	54.22%	22.24%
2250	1.0047	0.9016	53.78%	22.30%
2500	0.8928	0.8975	53.38%	22.12%

The best performance (lowest WER of 53.38% and lowest CER of 22.12%) is reached at the final step (2,500).

Qualitative Comparison (Before / After Fine-Tuning)

Reference sentence (expected transcription): "bon nalonaka bandunda lokola épinard"

—Before fine-tuning: the base model generates a French sentence entirely unrelated to the Lingala reference.
—After fine-tuning: the model produces "Bonalonaka bandunda lokola epinard", showing clearly improved language orientation and transcription fidelity.

Framework versions

—PEFT 0.20.0

Deployment Specification (API)

Component	Detail
Inference server	vLLM with required audio dependencies
Endpoint	`POST ${VLLM_BASE_URL}/v1/audio/transcriptions`
Request type	`multipart/form-data`
Inputs	Audio file, model identifier, and optional parameters (`language`, `prompt`, `response_format`, `temperature`)
Accepted audio formats	MP3, MP4, M4A, WAV, WEBM, OGG, FLAC, MPEG, MPGA
Output	Raw text transcription or detailed JSON response (including detected language, duration, and time-aligned segments)
LiteLLM alias	Normalized to the public name `asr` (with `asr-lingala` and `asrl` temporarily kept for historical compatibility)

Validation Status and Evidence Limitations

Although the merged model is available on Hugging Face, the application-level endpoint remains "to be deployed and validated" — pending a stable production URL, an operational authentication key, and real functional testing.

The exact behavior of the ASR pipeline through LiteLLM routing and the final inference provider still needs to be validated through real end-to-end audio inference tests.

Bias, Risks, and Limitations

This model was fine-tuned on an internal Lingala speech corpus of limited scope. Word Error Rate remains relatively high (53.38%), so transcriptions should be treated as a draft requiring human review rather than a final, publication-ready output — particularly for regional accents, dialectal variation, or noisy audio conditions not well represented in the training data.

Environmental Impact

Carbon emissions can be estimated using the Machine Learning Impact calculator (Lacoste et al., 2019).

Citation

BibTeX:

bibtex

@misc{cds2026whisperlingala,
  title={whisper-small-lingala-qlora},
  author={Congo Digital Services},
  year={2026},
  howpublished={\url{https://huggingface.co/Congo-digital-service/whisper-small-lingala-qlora-merged}}
}

Contact

—Organization: Congo Digital Services (CDS SARL)
—Website: congo-digital.com
—Email: info@congo-digital.com