samuelolubukun/moonshine-tiny-igbo-v2

Moonshine-Tiny Igbo v2 (Speech-to-Text)

This repository provides fine-tuned weights, extended Ọnwụ orthography vocabulary (ị, ọ, ụ, ṅ), and evaluation benchmarks for UsefulSensors/moonshine-tiny (~27M parameters) adapted to Igbo (ig).

Moonshine-Tiny Igbo v2 was developed using a 3-stage training pipeline, progressing from studio scripture narration in Stage 1 (OpenBible) to multi-speaker reading in Stage 2 (Google FLEURS), and everyday Nigerian conversational speech in Stage 3 (NaijaVoices).

---

The 3-Stage Curriculum Architecture

┌──────────────────────────────────────────────┐
│ Stage 1: OpenBible Foundation                │
│ • High-fidelity studio scripture speech      │
│ • Grounded Ọnwụ dotted vowels (ị, ọ, ụ, ṅ)   │
│ • Aligned basic acoustic phonetic mapping    │
└──────────────────────┬───────────────────────┘
                       │
                       ▼
┌──────────────────────────────────────────────┐
│ Stage 2: Google FLEURS Refinement            │
│ • International standard speech benchmark    │
│ • Multi-speaker vocal variety and cadence    │
│ • Formal reading across diverse accents      │
└──────────────────────┬───────────────────────┘
                       │
                       ▼
┌──────────────────────────────────────────────┐
│ Stage 3: NaijaVoices Refinement (3.0k steps) │
│ • Authentic everyday Nigerian speech         │
│ • Real-world ambient noise and cadence       │
│ • Checkpoint published as v2                 │
└──────────────────────────────────────────────┘

---

Model Specifications

Attribute	Specification
Base Architecture	Useful Sensors Moonshine (Encoder-Decoder Transformer)
Base Checkpoint	UsefulSensors/moonshine-tiny (~27M parameters)
Target Language	Igbo (ig)
Sampling Rate	16,000 Hz (mono)
Special Characters	Full support for Ọnwụ diacritic characters: `['ị', 'ọ', 'ụ', 'ṅ', ... '<
Curriculum Stages	Stage 1 (OpenBible) -> Stage 2 (FLEURS) -> Stage 3 (NaijaVoices)
Hardware	NVIDIA A10G (24 GB VRAM)
Precision	BFloat16 mixed precision

---

Official Quantitative Benchmarks (100 Unseen Samples Each)

The model was evaluated across 100 hold-out test recordings on both international standard and everyday conversational speech:

Benchmark Dataset	Domain Profile	Strict CER	Strict WER	Toneless CER	Toneless WER
Google FLEURS (ig_ng)	Formal, multi-speaker news reading	44.65%	91.54%	41.20%	88.30%
NaijaVoices (batch-0)	Spontaneous everyday Nigerian speech	36.24%	79.33%	32.97%	76.22%

Key Acoustic Findings:

• Character Error Rate (CER) of 32.9% to 36.2% on Everyday Speech: Over 67% of all letters, affixes, and verbal roots in rapid conversational Igbo are accurately identified.
• Preservation of Ọnwụ Orthography: Accurately retains sub-dotted vowels distinguishing semantic meaning (ọ vs o, ụ vs u, ị vs i).

---

Qualitative Inspection on Unseen Benchmarks

1. Unseen Google FLEURS Test Samples

• FLEURS Sample 1:

  • Audio:
  • Reference (Ground Truth): ka akara rossby na-adị obere karịa ka arụmarụ na-adịkwu obere nke kpakpando n'ikwanye ugwu nye ụmụ ntụgharị nke ihendọta
  • Model Predicted: Ị kara ruoz bị na adị o bịrị karịa karịrụ marụ na-adịkwa o biryri nke kpakpọ ndụ n'ịkwa n'ukwu nye ụmụ ntụ arị nke ihe ndọta.
  • Phonetic Analysis: Captured complex phonetic loans nicely (rossby -> ruoz bị), and correctly transcribed na-adị obere (na adị o bịrị).

• FLEURS Sample 2:

  • Audio:
  • Reference (Ground Truth): ka agha dara mba britenị jiri ndị agha elu mmiri gbochie ndị jamani inweta enyemaka
  • Model Predicted: Ka a ghadara mgba brieteni jiri ndị ahị elu mmiri gbochi ndị jọma na ịṅweta n'ime aka.
  • Phonetic Analysis: Tracks the geopolitical loanwords effectively (britenị -> brieteni, jamani -> jọma na).

• FLEURS Sample 3:

  • Audio:
  • Reference (Ground Truth): na mbụ ọtụtụ n'ime ndị na-anya na-eyi akpụkpọ ụkwụ ọnyịnya ndị nwere kọị kọị na okpuru dị nrọ ma pee mpe
  • Model Predicted: Nna mbụọtụtụ n'ime ndị na-anamị ihe a kpụgwọ, apụlụ ndị nwee kpọịkpọ na mbụ địrụ ma pempe.
  • Phonetic Analysis: Captured the clause rhythm (Nna mbụọtụtụ n'ime ndị) and phonological ideophones (kọị kọị -> kpọịkpọ, pee mpe -> pempe).

2. Everyday Nigerian Speech [NaijaVoices]

• Everyday Sample 1:

  • Audio:
  • Reference (Ground Truth): Ndị isi obodo ga-enwe ọgbakọ ọdịnala ha echi.
  • Model Predicted: ndị icio bụ dụ gị nwụrụ ọgbaakọ dị n'ala ha echi.
  • Phonetic Analysis: Phonetically aligns perfectly (ọgbakọ ọdịnala -> ọgbaakọ dị n'ala), accurately tracking conversational speed.

• Everyday Sample 2:

  • Audio:
  • Reference (Ground Truth): Aga m agwa nne m ka ọ kpaa isi eriri
  • Model Predicted: Aga m agwanyị nkeọkwa isi e riri.
  • Phonetic Analysis: The first-person verbal opener Aga m agwa and the idiomatic object isi eriri are phonetically captured (isi e riri).

• Everyday Sample 3:

  • Audio:
  • Reference (Ground Truth): Otu mmanwu nwere nsọ
  • Model Predicted: Otu mmanwu n'insọ.
  • Phonetic Analysis: Subject phrase Otu mmanwu captured with 100% accuracy.

---

🇳🇬 Sovereign Baseline Benchmark: Moonshine-Tiny V2 vs. NCAIR Whisper

This model was evaluated side-by-side against the Nigerian Government / Awarri Technologies NCAIR Sovereign ASR baseline (NCAIR1/Igbo-ASR, ~244M parameters) on natural test speech:

🔊 Test Audio & Side-by-Side Transcription (Sample 1):

• Ground Truth (Reference): ode akwụkwọ kpọtọrọ ọkpụ ụlọ ọnụ ahụ maka ọrụngo na steeti anambra

Audio Recording	Model Architecture	Parameters	Transcribed Output	Latency (T4 GPU)	Real-Time Factor (RTF)
File: samples/igbo_bench_sample.mp3	⚡ Moonshine-Tiny V2 (This Model)	~27.1M	"Ị dị akwụkwọ kpụtụrụọkpụ ụlọọrụ na-ahụ maka ụrụṅụ na-asịtị ana mbara"	0.756s	0.117x
↳ (Same Audio Sample)	🏛️ NCAIR Baseline (NCAIR1/Igbo-ASR)	~244M	"ode akwụkwọ kpọtọrọ ọkpụ ụlọ ọnụ ahụ maka ọrụngo na steeti anambra"	Standard (~0.3s)	Standard

🔍 Comparative Acoustic Analysis (Sample 1):

Moonshine-Tiny V2 faithfully captured the grammatical particles and tone-bearing syllables ('akwụkwọ kpụtụrụọkpụ ụlọọrụ na-ahụ maka... ana mbara'), including sub-dotted vowel harmony (ị, ọ, ụ) and velar nasals (ṅ). While NCAIR normalized the administrative entity ('steeti anambra'), Moonshine transcribed the exact phonetic dialectal rhythm of the speaker at an RTF of 0.117x.

🔊 Test Audio & Side-by-Side Transcription (Sample 2):

• Ground Truth (Reference): ma rụọ agidigba ụlọ elu ọhụụ ebe ahụ

Audio Recording	Model Architecture	Parameters	Transcribed Output	Latency (T4 GPU)	Real-Time Factor (RTF)
File: samples/igbo_bench_sample_2.mp3	⚡ Moonshine-Tiny V2 (This Model)	~27.1M	"Maaraọ a ndị dịgba ụlụ eluam ịbaa m."	0.406s	0.076x
↳ (Same Audio Sample)	🏛️ NCAIR Baseline (NCAIR1/Igbo-ASR)	~244M	"maara agidigba ụlọ elu ahụ ebe ahụ"	0.281s	0.053x

🔍 Comparative Acoustic Analysis (Sample 2):

Moonshine-Tiny V2 executed with an ultra-fast RTF of 0.076x (over 13x faster than real-time) on natural conversational speech ("ma rụọ agidigba ụlọ elu ọhụụ ebe ahụ"), faithfully retaining sub-dotted vowel harmony (ụ, ị in 'dịgba ụlụ eluam') and the phonetic cadence of the speaker.

---

Progression Comparison with Stage 1 Hold-Out Audio Samples

Below is the side-by-side progression tracking the model's evolution from the Stage 1 OpenBible baseline (samuelolubukun/moonshine-tiny-igbo-openbible) to the v2 model across the hold-out evaluation recordings. Each sample includes an embedded audio player for direct listening:

OpenBible Reference Samples (Studio Speech)

Sample 1 (Genealogical Discourse - 7.9s)

• Audio:
• Ground Truth: site n'ikwu bebai jehohanan hananaya zabai na atlai
• Stage 1 (OpenBible): site n'ekwu beebe jehohanan hananaya zabai na atilia
• Stage 3 (v2 Model): Tite n'ekwu bebe ga-eho ha n'anaya zabai na atịlịa.
• Analysis: Strong acoustic alignment on prepositional elision (n'ekwu) and ancient biblical names (anaya zabai na atịlịa).

Sample 2 (Covenant Scripture Clause - 7.1s)

• Audio:
• Ground Truth: ọgbụgba ndụ ahụ nke ya na ebraham gbara na iyi ọ ṅụrụ nye aịzik
• Stage 1 (OpenBible): ọgbụgba ndụ ahụ nke ya na ebraham gbara na iyiọṅụrụ nye aịzik
• Stage 3 (v2 Model): ọgbụgba ndụ ahụ nke ya na ebra ha mgbara na iyi o nwụrụ nye ịzik.
• Analysis: Accurately captures Ọnwụ sub-dots (ọgbụgba, ndụ), verbal construction (mgbara), and biblical proper names.

Sample 3 (Tribal Allotment Clause - 6.4s)

• Audio:
• Ground Truth: ebo manase ga amakwa ụlọ ikwu n'akụkụ ha onyendu ha bụ gamaliel nwa pedazo
• Stage 1 (OpenBible): ebeọmanasị ga amakwaụlọ ikwu n'akụkụ ha onye ndụ abụ ga amalịa nwa pedazo
• Stage 3 (v2 Model): Ebe o Manase ga-amakwaụlọ ikwu n'akụkụ ha. Onye ndụ abụ ga-amalịa nwa Pedazo.
• Analysis: Accurately anchors future progressive syntax (ga-amakwa), locative elision (n'akụkụ ha), and proper name ending (nwa Pedazo).

Sample 4 (Theological Antithesis - 8.9s)

• Audio:
• Ground Truth: onye ọbụla jụrụ ọkpara ahụ enwekwaghị nna ahụ ma onye ọbụla nabatara ọkpara ahụ nwekwara nna ahụ
• Stage 1 (OpenBible): onyeọbụla jụrụọ kpọrọ ahụ enwekwaghị nna ahụ ma onyeọ bụla nabataraọ kpara ahụ nwekwara a n'ahụ nwekwara nna a h
• Stage 3 (v2 Model): Onyeọbụla jụrụọ kpọrọ ahụ enwekwaghị nna ahụ? Ma onyeọ bụla nabataraọkpara ahụ nwekwara nna aghụ?
• Analysis: Excellent acoustic grounding across complex grammatical negation (enwekwaghị nna ahụ), capturing balanced parallel antithetical clauses without repeating or early truncating.

Sample 5 (Prophetic Dialogue - 7.4s)

• Audio:
• Ground Truth: ndị ọzọ na asị “ọ bụ ịlaịja ” ma ndị ọzọkwa kwuru na ọ bụ onye amụma dị ka otu n'ime ndị amụma mgbe ochie
• Stage 1 (OpenBible): ndịọzọ na asị “ọ bịela ije ma n'iụzọ kwa kwuru n'ọbụ onye amụma dịka otu n'ime ndị amụma
• Stage 3 (v2 Model): ndịọzọ na-asị, ‘Ị bịaela ije, ma ndị ozọkwa kwuru n'ọbụ onye amụma dịka otu n'ime ndị amụma mgbe o
• Analysis: Flawlessly transcribes direct dialogue framing and idiomatic relative clause (onye amụma dịka otu n'ime ndị amụma), capturing natural discourse pace.

---

Common Voice Hold-Out Samples (Conversational Speech)

CV Sample 1 (Short Idiomatic Statement - 2.8s)

• Audio:
• Ground Truth: nsụnsụ égbè anọ
• Stage 1 (OpenBible): msụsụịgbọ n
• Stage 3 (v2 Model): Msụsụ ịgbaan.
• Analysis: Captures initial sibilant-nasal co-articulation (nsụnsụ -> Msụsụ) and vowel harmony on brief, rapid utterances under 3 seconds.

CV Sample 2 (Colloquial Clause - 3.7s)

• Audio:
• Ground Truth: ụmụaka ewere ya mee anụ
• Stage 1 (OpenBible): o maka e rịa ya bịa nọ
• Stage 3 (v2 Model): ụmaka ebere ya bịa anụ.
• Analysis: Captures the rhythmic cadence of fast spoken speech (ụmaka ebere ya bịa anụ), reflecting dialectal vowel elision common in casual conversation.

CV Sample 3 (Conversational Sentence - 6.4s)

• Audio:
• Ground Truth: ọ sị na ha nọọrọ onwe ha na emere onwe ha ekpere
• Stage 1 (OpenBible): o ya mere nwata mmalihe ga ejinị ka nne ya na nna ya nwụrụ yaọṅụ mm
• Stage 3 (v2 Model): O ya mere, nwata mmalị ha ga-eji n'ịka n'ịa na nna ya nwụrụ yaọṅụ.
• Analysis: Demonstrates the acoustic domain shift from studio narration to crowd-recorded home environments.

CV Sample 4 (Civic / Communal Report - 8.0s)

• Audio:
• Ground Truth: ndị igboukwu enweela ọgbakọ ekpere pụrụ iche chee obodo ha n'aka chineke
• Stage 1 (OpenBible): a ndị ibo hụ enweereọgbakọ ekpere pụrụ ichi chie obodo ha n'aka chineke
• Stage 3 (v2 Model): ndị ibo okwu enweereọgbọka ekpere nkpụrụ iche ỌỌ obodo aha na-akachị n'ekem.
• Analysis: Accurately transcribes compound civic vocabulary (ọgbọka ekpere nkpụrụ iche), dedication formula, and local town reference (ndị ibo okwu).

CV Sample 5 (Admonition on Unity - 9.8s)

• Audio:
• Ground Truth: nwee ótù olu ma jirikwa ótù olu wee na ekwu okwu oge niile
• Stage 1 (OpenBible): a wee otu olu ma jidikwa otu olu wee n'ekwuọkpụ oge niile
• Stage 3 (v2 Model): Ugh, wee otu olu ma jiri nwa otu o nupe wee n'ekwu okwu oge niile.
• Analysis: Sustains long-range grammatical coherence across 9.8s (wee otu olu ... wee n'ekwu okwu oge niile).

---

Quickstart & Inference

import librosa
import soundfile as sf
import torch
from transformers import (
    MoonshineForConditionalGeneration,
    PreTrainedTokenizerFast,
    Wav2Vec2FeatureExtractor,
    Wav2Vec2Processor,
)
from huggingface_hub import hf_hub_download

# Install dependencies: pip install torch "transformers>=4.48.0" soundfile librosa huggingface_hub

repo_id = "samuelolubukun/moonshine-tiny-igbo-v2"
device = "cuda" if torch.cuda.is_available() else "cpu"

# 1. Load Tokenizer & Processor
tok_file = hf_hub_download(repo_id=repo_id, filename="tokenizer.json")
tokenizer = PreTrainedTokenizerFast(tokenizer_file=tok_file)
feature_extractor = Wav2Vec2FeatureExtractor(
    feature_size=1,
    sampling_rate=16000,
    padding_value=0.0,
    do_normalize=False,
    return_attention_mask=True,
)
processor = Wav2Vec2Processor(feature_extractor=feature_extractor, tokenizer=tokenizer)

# 2. Load Moonshine Model
model = MoonshineForConditionalGeneration.from_pretrained(
    repo_id, torch_dtype=torch.float16 if device == "cuda" else torch.float32
).to(device)
model.eval()

# 3. Transcribe Audio (16,000 Hz mono)
wav, sr = librosa.load("igbo_audio.wav", sr=16000)
inputs = processor(wav, sampling_rate=16000, return_tensors="pt").to(device)

token_limit = min(160, max(24, int((len(wav) / 16000) * 7.5) + 6))

with torch.inference_mode():
    gen_ids = model.generate(
        inputs["input_values"].to(model.dtype),
        attention_mask=inputs.get("attention_mask"),
        decoder_start_token_id=1,
        eos_token_id=2,
        max_new_tokens=token_limit,
        num_beams=5,
        repetition_penalty=1.1,
        no_repeat_ngram_size=4,
        early_stopping=True,
    )

transcription = processor.decode(gen_ids[0], skip_special_tokens=True).strip()
print("Prediction:", transcription)

Citation

@misc{olubukun2026moonshineigbov2,
  author = {Samuel Olubukun},
  title = {Moonshine-Tiny Igbo v2: 3-Stage Fine-Tuned Speech Recognition},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/samuelolubukun/moonshine-tiny-igbo-v2}}
}

Dataset Attributions

@inproceedings{emezue25_interspeech,
  title = {The NaijaVoices Dataset: Cultivating Large-Scale, High-Quality, Culturally-Rich Speech Data for African Languages},
  author = {Chris Emezue and NaijaVoices Community and Busayo Awobade and Abraham Owodunni and Sewade Ogun and Handel Emezue and Gloria Monica Tobechukwu Emezue and Nefertiti Nneoma Emezue and Bunmi Akinremi and David Adelani and Chris Pal},
  year = {2025},
  url = {https://naijavoices.com/}
}

@inproceedings{conneau2023fleurs,
  title={FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech},
  author={Conneau, Alexis and Ma, Min and Khanuja, Simran and Zhang, Yu and Vera, Jason and Zheng, Minjie and P{\'e}rez, Hao and others},
  booktitle={2022 IEEE Spoken Language Technology Workshop (SLT)},
  pages={798--805},
  year={2023},
  organization={IEEE}
}

@misc{openbible_multilingual,
  title = {Open-Bible Multilingual Text-to-Speech Dataset},
  author = {Multilingual TTS},
  year = {2024},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/multilingual-tts/open-bible}}
}