PrinceAlhassanNasamu

kusaal-whisper-small-lora

README

License: other

Results

Table with columns: Split, Books held out, WER
Split	Books held out	WER
Val	LUK, PHP, PRO, PSA	~30.4%

Evaluated on 400 randomly sampled clips from the val split of alhassanprince/kusaal-asr-dataset. Test-set WER (JON, MAT, REV, RUT) to be reported in the accompanying paper.

Context: Kusaal had no prior ASR system or public speech corpus before this work. A 30.4% WER on a first model trained on ~82 hours of scripture audio, with book-held-out evaluation, represents a strong baseline for a zero-resource language.

Sample outputs

markdown
PSA.89.22
REF: dataa kae na nyaŋ oo bugusim sɔb kae na mugus oo
HYP: m taa kae na nyaŋ o bugusim sɔb kae na mugus o

PRO.26.15
REF: gbanyaa na nɔk o nuugu tɔm laan amaa ka ukim niŋ nɔɔrin la an yɛl
HYP: yaa ya na nɔk o nuugu tɔm laan amaa ka o ku m niŋ nɔɔrin la an yɛl

PSA.98.5
REF: nwɛɛm kɔnne tis zugsɔb la yɔɔm ka nwɛ kɔn
HYP: nwɛɛm kɔn nɛ tis zugsɔb la yɔɔm ka nwɛɛ kɔɔn

The model correctly handles Kusaal-specific characters: ŋ ɔ ɛ ʋ.

Training details

Table with columns: Parameter, Value
Parameter	Value
Base model	openai/whisper-small (241M params)
Method	LoRA (PEFT)
LoRA rank	32
LoRA alpha	64
LoRA targets	`q_proj`, `v_proj`
Trainable params	3,538,944 (1.44% of total)
Epochs	2
Optimizer steps	3,074

Training data

Trained on alhassanprince/kusaal-asr-dataset:

30,820 verse-level WAV clips, ~81.7 hours
16 kHz mono 16-bit PCM WAV
Kusaal orthographic transcripts (lowercase; includes ŋ ɔ ɛ ʋ)
Book-held-out splits across 63 books of the Bible
Train split: 24,597 clips · 69.33 hrs · 55 books

Usage

Load and transcribe

python
import torch
import soundfile as sf
from transformers import WhisperProcessor, WhisperForConditionalGeneration
from peft import PeftModel

# Load base model + LoRA adapter
base = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")
model = PeftModel.from_pretrained(base, "PrinceAlhassanNasamu/kusaal-whisper-small-lora")
processor = WhisperProcessor.from_pretrained("openai/whisper-small")

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

# Transcribe a WAV file (must be 16 kHz mono)
audio, sr = sf.read("your_audio.wav", dtype="float32")
inputs = processor.feature_extractor(audio, sampling_rate=16000, return_tensors="pt")
input_features = inputs.input_features.to(device)

with torch.no_grad():
    ids = model.generate(
        input_features=input_features,
        max_length=160,
        no_repeat_ngram_size=3,
    )

transcript = processor.batch_decode(ids, skip_special_tokens=True)[0]
print(transcript)

Requirements

bash
pip install transformers peft soundfile torch

Limitations

Single domain — trained on scripture audio only; may underperform on conversational, broadcast, or spontaneous Kusaal speech
Single speaker style — limited speaker diversity in training data
Book-domain gap — val/test books (LUK, PSA, MAT, REV etc.) represent a distribution shift; conversational speech will be larger
Orthography — transcripts are lowercase normalised; punctuation limited
Repetition — rare hallucination loops on difficult clips (mitigated with no_repeat_ngram_size=3)

Audio sources & acknowledgments

The speech audio underlying this model's training data comes from scripture recording ministries who record and distribute audio in minority languages as a free public service. Their work made this dataset possible.

Faith Comes By Hearing (FCBH) / Hosanna

New Testament audio: Kusaal NT Drama recording (KUSTBLN2DA)
© Wycliffe / Hosanna (1996, 2005); distributed by Faith Comes By Hearing
faithcomesbyhearing.com

Davar Partners International

Old Testament audio: Kusaal OT recording
© Davar Partners International (2023)
davaraudiobibles.org

GILLBT — Ghana Institute of Linguistics, Literacy and Bible Translation

Kusaal Bible text (GILLBT translation) used for all transcript alignment
gillbt.org

Global Recordings Network (GRN)

Additional audio: Kusal "Look, Listen & Live" series and "Words of Life" programs
globalrecordings.net

Please respect their terms when using this model: non-commercial and research/education use only. Do not redistribute the underlying audio commercially without clearing rights with the original rights-holders directly.

License

This model is a derivative work with multiple licensing layers:

Table with columns: Layer, Source, Terms
Layer	Source	Terms
Base model weights	openai/whisper-small	MIT
LoRA adapter weights	This repo	CC BY 4.0 (research use)
Training audio	FCBH, Davar, GRN	Non-commercial only
Training text	GILLBT Kusaal Bible	Non-commercial only

For commercial deployment, contact the original audio rights-holders:

Table with columns: Resource, Link
Resource	Link
ASR dataset	alhassanprince/kusaal-asr-dataset — 81.7h, 30,820 clips
MT corpus	PrinceAlhassanNasamu/kusaal-english-parallel-corpus — 34,568 parallel pairs
MT model	PrinceAlhassanNasamu/kusaal-nllb-600M — NLLB-200 fine-tune, BLEU 27.57 kus→eng
GhanaNLP	ghananlp.org

Citation

bibtex
@misc{alhassanprince_kusaal_asr_2026,
  author    = {Alhassan, Prince Nasamu},
  title     = {Kusaal Whisper-small LoRA: First ASR Model for Kusaal},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/PrinceAlhassanNasamu/kusaal-whisper-small-lora},
  note      = {LoRA fine-tune of openai/whisper-small; VAL WER ~30.4\%; 
               trained on 81.7h of verse-level Kusaal scripture audio}
}

Author

Prince Nasamu Alhassan
Independent NLP Researcher · University of Ghana, Legon · GhanaNLP
HuggingFace: PrinceAlhassanNasamu
Kaggle: alhassanprince

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

PrinceAlhassanNasamu

Model Tree

Base

openai/whisper-small

Adapter

this model

Input Modalities

Audio

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Container

Explore FriendliAI today

Get started Talk to an engineer

README

License: other

Results

Table with columns: Split, Books held out, WER
Split	Books held out	WER
Val	LUK, PHP, PRO, PSA	~30.4%

Evaluated on 400 randomly sampled clips from the val split of alhassanprince/kusaal-asr-dataset. Test-set WER (JON, MAT, REV, RUT) to be reported in the accompanying paper.

Context: Kusaal had no prior ASR system or public speech corpus before this work. A 30.4% WER on a first model trained on ~82 hours of scripture audio, with book-held-out evaluation, represents a strong baseline for a zero-resource language.

Sample outputs

markdown
PSA.89.22
REF: dataa kae na nyaŋ oo bugusim sɔb kae na mugus oo
HYP: m taa kae na nyaŋ o bugusim sɔb kae na mugus o

PRO.26.15
REF: gbanyaa na nɔk o nuugu tɔm laan amaa ka ukim niŋ nɔɔrin la an yɛl
HYP: yaa ya na nɔk o nuugu tɔm laan amaa ka o ku m niŋ nɔɔrin la an yɛl

PSA.98.5
REF: nwɛɛm kɔnne tis zugsɔb la yɔɔm ka nwɛ kɔn
HYP: nwɛɛm kɔn nɛ tis zugsɔb la yɔɔm ka nwɛɛ kɔɔn

The model correctly handles Kusaal-specific characters: ŋ ɔ ɛ ʋ.

Training details

Table with columns: Parameter, Value
Parameter	Value
Base model	openai/whisper-small (241M params)
Method	LoRA (PEFT)
LoRA rank	32
LoRA alpha	64
LoRA targets	`q_proj`, `v_proj`
Trainable params	3,538,944 (1.44% of total)
Epochs	2
Optimizer steps	3,074

Training data

Trained on alhassanprince/kusaal-asr-dataset:

30,820 verse-level WAV clips, ~81.7 hours
16 kHz mono 16-bit PCM WAV
Kusaal orthographic transcripts (lowercase; includes ŋ ɔ ɛ ʋ)
Book-held-out splits across 63 books of the Bible
Train split: 24,597 clips · 69.33 hrs · 55 books

Usage

Load and transcribe

python
import torch
import soundfile as sf
from transformers import WhisperProcessor, WhisperForConditionalGeneration
from peft import PeftModel

# Load base model + LoRA adapter
base = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")
model = PeftModel.from_pretrained(base, "PrinceAlhassanNasamu/kusaal-whisper-small-lora")
processor = WhisperProcessor.from_pretrained("openai/whisper-small")

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

# Transcribe a WAV file (must be 16 kHz mono)
audio, sr = sf.read("your_audio.wav", dtype="float32")
inputs = processor.feature_extractor(audio, sampling_rate=16000, return_tensors="pt")
input_features = inputs.input_features.to(device)

with torch.no_grad():
    ids = model.generate(
        input_features=input_features,
        max_length=160,
        no_repeat_ngram_size=3,
    )

transcript = processor.batch_decode(ids, skip_special_tokens=True)[0]
print(transcript)

Requirements

bash
pip install transformers peft soundfile torch

Limitations

Single domain — trained on scripture audio only; may underperform on conversational, broadcast, or spontaneous Kusaal speech
Single speaker style — limited speaker diversity in training data
Book-domain gap — val/test books (LUK, PSA, MAT, REV etc.) represent a distribution shift; conversational speech will be larger
Orthography — transcripts are lowercase normalised; punctuation limited
Repetition — rare hallucination loops on difficult clips (mitigated with no_repeat_ngram_size=3)

Audio sources & acknowledgments

Faith Comes By Hearing (FCBH) / Hosanna

New Testament audio: Kusaal NT Drama recording (KUSTBLN2DA)
faithcomesbyhearing.com

Davar Partners International

Old Testament audio: Kusaal OT recording
davaraudiobibles.org

GILLBT — Ghana Institute of Linguistics, Literacy and Bible Translation

Kusaal Bible text (GILLBT translation) used for all transcript alignment
gillbt.org

Global Recordings Network (GRN)

Additional audio: Kusal "Look, Listen & Live" series and "Words of Life" programs
globalrecordings.net

License

This model is a derivative work with multiple licensing layers:

Table with columns: Layer, Source, Terms
Layer	Source	Terms
Base model weights	openai/whisper-small	MIT
LoRA adapter weights	This repo	CC BY 4.0 (research use)
Training audio	FCBH, Davar, GRN	Non-commercial only
Training text	GILLBT Kusaal Bible	Non-commercial only

For commercial deployment, contact the original audio rights-holders:

Table with columns: Resource, Link
Resource	Link
ASR dataset	alhassanprince/kusaal-asr-dataset — 81.7h, 30,820 clips
MT corpus	PrinceAlhassanNasamu/kusaal-english-parallel-corpus — 34,568 parallel pairs
MT model	PrinceAlhassanNasamu/kusaal-nllb-600M — NLLB-200 fine-tune, BLEU 27.57 kus→eng
GhanaNLP	ghananlp.org

Citation

bibtex
@misc{alhassanprince_kusaal_asr_2026,
  author    = {Alhassan, Prince Nasamu},
  title     = {Kusaal Whisper-small LoRA: First ASR Model for Kusaal},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/PrinceAlhassanNasamu/kusaal-whisper-small-lora},
  note      = {LoRA fine-tune of openai/whisper-small; VAL WER ~30.4\%; 
               trained on 81.7h of verse-level Kusaal scripture audio}
}

Author

Prince Nasamu Alhassan
Independent NLP Researcher · University of Ghana, Legon · GhanaNLP
HuggingFace: PrinceAlhassanNasamu
Kaggle: alhassanprince

kusaal-whisper-small-lora

README

Results

Sample outputs

Training details

Training data

Usage

Load and transcribe

Requirements

Limitations

Audio sources & acknowledgments

License

Related work

Citation

Author

Explore FriendliAI today

README

Results

Sample outputs

Training details

Training data

Usage

Load and transcribe

Requirements

Limitations

Audio sources & acknowledgments

License

Related work

Citation

Author