Introduction

Sema TTS Boost is the same-week unlock for makers who already eyed Sema’s African F5 stack and wanted the weak languages louder. On 2026-11-04 (~19:29 UTC) Angelo Asante published travis-moore/sema-tts-boost — a targeted continuation of sema-tts-afro that keeps the 24-variety zero-shot voice-cloning surface and spends ~313 hours of extra audio on the languages that were lagging.

This is generative text-to-speech with cloning: a short reference clip sets the speaker; the model speaks new text in Twi, Hausa, Yoruba, Pidgin, Swahili, and the rest of the Sema map.

What Shipped

  • Vendor / product: Sema / Angelo Asante — Sema TTS Boost v1
  • Modality: open-weight multilingual text-to-speech + zero-shot voice cloning across 24 African language varieties (plus English / Pidgin)
  • Open vs closed: Apache-2.0 weights; F5-TTS stack (library_name: f5-tts), seeded from the base model’s EMA
  • Runnable today: pip install f5-tts==1.1.22, then hf_hub_download("travis-moore/sema-tts-boost", "sema_boost_v1.pt") + vocab.txt and F5TTS(...).infer(...)
  • Concrete fact: booster training adds ~213 h focus speech (Hausa, Yoruba, Amharic, Twi, West African Pidgin, multi-accent English) plus ~100 h replay; 4 epochs / 57,056 updates at 5e-6 LR; Hausa MMS CER gap improves +18.4 → +8.5 (Δ −9.9)

The card is honest about trade-offs: average CER gap across 21 scoreable languages moves +6.1 → +7.1 because a few non-focus languages drift, while listening tests (not ASR) are the stated reason Boost exists for Pidgin / Twi / accented English.

Run It

pip install f5-tts==1.1.22

from huggingface_hub import hf_hub_download
from f5_tts.api import F5TTS

ckpt = hf_hub_download("travis-moore/sema-tts-boost", "sema_boost_v1.pt")
vocab = hf_hub_download("travis-moore/sema-tts-boost", "vocab.txt")

tts = F5TTS(
    model="F5TTS_v1_Base",
    ckpt_file=ckpt,
    vocab_file=vocab,
    use_ema=True,
)
tts.infer(
    ref_file="speaker.wav",  # 5–10 s clean reference
    ref_text="what the speaker says in that clip",
    gen_text="Text to speak, in any supported language.",
    file_wave="out.wav",
    nfe_step=32,
    speed=0.9,
    seed=42,
)

Weights: sema_boost_v1.pt (~5.4 GB) + vocab.txt on the HF card. Same 3,821-token vocabulary as the base — drop-in compatible. With this extended vocab and f5-tts==1.1.22, the card asks you to patch padding_value=-1 → 0 in f5_tts/model/utils.py before batch inference.

Notes for makers

  • Prefer Boost when Hausa, Pidgin, Twi, accented English, Northern Sotho, or Fula matter; keep the base if Shona, Somali, or Wolof are the priority (Yoruba is roughly even)
  • Amharic still lags hard (+78.8 → +75.0): the card blames data quality, not more epochs
  • Clean, calmly paced 5–10 s references clone best; the model copies reference rhythm; prefer short sentences over comma chains
  • Base Sema Afro landed 2026-11-02 (outside this cycle’s ~48h cut) — Boost is the <48h primary
  • Distinct from ArtRealmAI’s Hubi Somali TTS and Ghana/Moore soft-holds — different vendor, F5 continuation, measured CER deltas
  • No Gen CTA: African F5 cloning does not map to the Flux / Grok / MiniMax-H3 / Seedream / Seedance Gen loop

Conclusion

If you need zero-shot African-language speech with a clearer Hausa / Pidgin / Twi path than the Nov 2 base, Sema TTS Boost v1 is today’s runnable drop: Apache-2.0 F5 weights, paste-ready f5-tts infer, and a documented −9.9 Hausa CER-gap win.

—Anabel ♡