Research & Model Checkpoints
Welcome to my ML research hub. My work explores applied machine learning across diverse frontiers, including but not limited to agentic systems, multi-modal reasoning, and compact on-device architectures. I plan to publish research and checkpoints across several machine learning sub-fields over time.
My current active focus is concentrated on sovereign speech technologies, domain adaptation, vocabulary transplantation, and regional dialect acoustic modeling. As of this writing, I have published 19 open-source fine-tuned speech models across Hugging Face spanning on-device ASR (Moonshine-Tiny), conversational neural voice cloning (XTTS-v2), and flow-matching synthesis (Chatterbox T3) across Yoruba (`yo`), Igbo (`ig`), Hausa (`ha`), Nigerian Pidgin (`pcm`), and Nigerian-Accented English (`en-ng`).
Note: This research page is continuously updated as new experiments and architectural benchmarks conclude. Keep an eye on my Hugging Face profile for the latest checkpoints, weights, and live releases.
XTTS-v2 Yoruba (OpenBible Conversational)
Conversational voice cloning adapted from the OpenBible foundation. Elevates speech tempo to 160-185 WPM with natural dynamic pitch glide and co-articulation.
XTTS-v2 Igbo (OpenBible Conversational)
Adaptive zero-shot cloning in Igbo with natural downstep intonation curves, shortened pauses, and fluent compound word blending.
XTTS-v2 Hausa (OpenBible Conversational)
Fast-tempo conversational Hausa voice synthesis (+38% speaking rate) with loudness normalization and robust hooked consonant preservation.
Moonshine-Tiny Hausa v2 (Multi-Domain)
Robust everyday speech recognition in Hausa adapting from scripture through broadcast news to crowdsourced mobile audio with zero hallucinations.
Moonshine-Tiny Yoruba v2 (Multi-Domain)
Refined 3-stage model transcribing colloquial Yoruba with phone mic acoustics and background noise while accurately decoding tonal diacritics.
Moonshine-Tiny Igbo v2 (Multi-Domain)
Multi-stage Igbo ASR adapting from high-SNR scripture to diverse accents and informal smartphone recordings with boundary healing.
Moonshine-Tiny Nigerian Pidgin
Ultra-fast on-device ASR for Nigerian Pidgin with extended SentencePiece dialect vocabulary, duration-scaled decoding, and 0.082x RTF on GPU.
Moonshine-Tiny Nigerian Accented English
High-throughput edge ASR for Nigerian-accented English running at 19.6x real-time speed (0.051x RTF) on GPU with sub-200ms processing latency.
Chatterbox T3 Yoruba (NaijaVoices)
Autoregressive token predictor with flow matching conditioned on Nigerian phone speech. Warm-started via Swahili embedding transplant with optimal stopping at Epoch 4.
Chatterbox T3 Igbo (NaijaVoices)
Zero-shot voice cloning with Encodec acoustic tokens and flow matching. Reaches optimal loss convergence at Epoch 4 before weight blending.
Chatterbox T3 Hausa (NaijaVoices)
High-expressivity Hausa speech synthesis trained on authentic Nigerian audio. Eliminates gradient spikes through cross-lingual token transfer.
Chatterbox T3 Nigerian Pidgin
Conversational zero-shot voice cloning for Nigerian Pidgin using English donor embedding transplantation, adaptive token budgeting, and calibrated VAD tail trimming.
Chatterbox T3 Nigerian Accented English
Flow-matching voice synthesis and cloning for Nigerian-accented English. Preserves authentic prosodic cadence and eliminates tail noise artifacts.
XTTS-v2 Yoruba (OpenBible Foundation)
Pristine studio acoustic foundation for Yoruba speech synthesis with extended 2,000-token BPE vocabulary for register tones.
XTTS-v2 Igbo (OpenBible Foundation)
High-SNR acoustic foundation for Igbo text-to-speech. Stabilizes articulation across dotted vowels (ị, ọ, ụ) and velar nasal ṅ.
XTTS-v2 Hausa (OpenBible Foundation)
Studio-grade Hausa foundation checkpoint preserving glottalized and implosive consonants (ɓ, ɗ, ƙ, ƴ) with zero phoneme degradation.
Moonshine-Tiny Igbo (OpenBible Foundation)
Compact on-device ASR model for Igbo with extended Ọnwụ orthography tokens (ị, ọ, ụ, ṅ). Clean phonetics with 0% citation hallucination.
Moonshine-Tiny Yoruba (OpenBible Foundation)
Edge encoder-decoder ASR with 29 expanded diacritic & tone units. Clean baseline phonetic acoustic model with boundary healing.
Moonshine-Tiny Hausa (OpenBible Foundation)
Compact Hausa speech recognizer natively encoding hooked consonants (ɓ, ɗ, ƙ, ƴ). Right-shifted BOS formatting eliminates looping.
Looking for the full engineering write-ups?
Read the detailed breakdowns covering training routines, tokenizer expansions, audio pre-processing, and benchmark evaluations across the models.