21M-param offline TTS model covers 8 African languages with 41 voices in just 40MB
strickvl · x · 2026-11-18
Developer samuelolubukun released Murivo-TTS Octa on Hugging Face: an ultra-lightweight, fully offline TTS engine with 21M parameters (40MB) covering 8 major African languages (Swahili, Hausa, Yorùbá, Igbo, Luganda, Fulfulde, Luo, Twi) with 41 voices.
- Architecture: 20.9M-param Flow-Matching DiT acoustic backbone (8 layers, RoPE Aligner) + VoxCPM2 48kHz Causal AudioVAE vocoder; native 48kHz studio audio at 13–21× real time on standard GPUs.
- Edge deployment: 15.5MB INT8-quantized ONNX version for fully offline use.
- Training: warm-started from an English 21M DiT backbone; data includes Google's WaxalNLP broadcast corpus, NaijaVoices and dedicated studio Yorùbá; stratified multilingual batching with per-speaker conditioning slots prevents timbre bleed across tonal/non-tonal languages; duration head has a 160ms floor.
- License: Apache-2.0.
More from Multimodal
- Claude Plus Unreal Engine 5.9 MCP Builds a Game Prototype in One Day — TomLikesRobots · 2026-11-18
- Suno's new Visuals feature pairs AI-generated music with AI-generated video — The Decoder · 2026-11-18
- AURA 2.0: a streaming video LLM that speaks first, calls tools, and grounds visually — Shuhan Li · 2026-11-18
- Pixel nun run cycle: extracting frames from a video model fixed the sliding — Reyan999 · 2026-11-18
- One mug, three photos: $4, $18 and $80 — an AI photo workflow for Black Friday pricing — LudovicCreator · 2026-11-18
- Midjourney style code --sref 3859897744 produces bold painterly illustration looks — sergeantsref · 2026-11-18