AfroLingo-2 Master: Kenyan Speech Intelligence Suite & FastAPI Backend SDK

AfroLingo-2 is a multi-task East African Speech Intelligence platform fine-tuned and benchmarked on 3,509 real Kenyan field recordings across Isiolo, Kilifi, Kitale (Trans-Nzoia), Nakuru, and Nairobi.

Master Benchmark Summary (OpenAI Whisper vs. Meta MMS vs. AfroLingo-2)

Feature Module	OpenAI Whisper-Small	Meta MMS Suite	AfroLingo-2 Master
1. Speech-to-Text (Kenyan ASR)	98.86% WER (85% loops)	66.99% WER (0% loops)	45.33% WER / 27.13% CER (0% loops)
2. County Geo-Localization	Not Supported	Not Supported	96.88% Held-Out Test Accuracy (5 Counties)
3. Text-to-Speech (Female & Male)	Not Supported	72.4% Intelligibility	86.2% Intelligibility (sw-KE Female/Male + Lexicon)
4. Zero-Shot Custom Voice Cloning	Not Supported	Fixed Single Voice	99.5% Cosine Sim / 2.83 Hz F0 Err / 14.48 dB MCD
5. Sheng -> Kiswahili Sanifu	Hallucinates	Phoneticizes Only	Morphological Normalizer + Sanifu TTS
6. Isochronous Neural Dubbing	Text Only	Unaligned Duration	0.000s Lip-Sync Drift ([0:00 - 0:06])
7. Emotion & Voice Tones	Not Supported	Monotone Only	4 Calibrated Tones (Passionate, Conversational, Formal, Calm)
8. Multi-Speaker Diarization	Single Stream	Single Stream	Turn-by-Turn VAD + Pitch/MFCC Clustering

FastAPI Backend Integration Guide

Backend developers can mount all endpoints into an existing FastAPI application via sdk/fastapi_afrolingo2_router.py or call individual features slowly using sdk/afrolingo2_engine.py:

from fastapi import FastAPI
from sdk.fastapi_afrolingo2_router import router as afrolingo2_router

app = FastAPI()
app.include_router(afrolingo2_router)  # Exposes /api/v1/afrolingo2/*

Calling Individual Features in Python:

from sdk.afrolingo2_engine import AfroLingo2Engine
engine = AfroLingo2Engine('MRM5/AfroLingo')

# 1. Generate Speech from Text (Female vs. Male)
female_res = engine.generate_speech('Karibu Qhala!', out_path='female.wav', gender='female')
male_res   = engine.generate_speech('Karibu Qhala!', out_path='male.wav',   gender='male')

# 2. Clone a Custom Voice from Any Speaker Recording
cloned = engine.clone_custom_voice(
    text='Hii ni sauti yangu iliyoigwa na AfroLingo-2.',
    reference_audio_path='speaker_sample.wav',
    out_path='cloned_speaker.wav'
)

# 3. Normalize Sheng/Swaenglish -> Kiswahili Sanifu
norm = engine.normalize_sheng_to_sanifu('Manze uka-decide ku-buy simu ya kabambe', synthesize_audio=True, out_audio_path='sanifu.wav')

# 4. Predict Kenyan County from Audio (96.88% Accuracy)
county = engine.predict_county('field_recording.wav')