travis-moore/sema-sauti-v1

Sema Sauti v1

Sema's Ghanaian voice: one cloned voice that speaks Twi, Nigerian and Ghanaian Pidgin, and English in a single model. Give it a few seconds of reference audio and it speaks all three in that voice.

Built under Sema by Angelo Asante.

Sema Sauti v1 is the recommended build: the lowest error on our read-back test and the point where the voice reads most naturally.

Hear it

Reference-cloned voices (exaggeration=0.5, cfg_weight=0.5).

Clip	What it is
	Twi, author's voice: “Me da wo ase paa, Onyankopɔn nhyira wo.”
	Twi–English code-switch: “Ɛnnɛ deɛ the meeting go start at nine o'clock.”
	Nigerian Pidgin, Nigerian voice: “Abeg make you no vex, I go explain everything.”
	Ghanaian Pidgin, author's voice: “Chale, the game yesterday be mad, you for see am.”
	English, author's voice.
	French — one of the languages the model also speaks.
	Swahili — another language it carries.

Languages

Three are trained targets — English, Twi/Akan, and Nigerian + Ghanaian Pidgin. Alongside them the model speaks a set of other languages it carried through training. We checked the whole set with a read-back test: one sentence per language, transcribed by Whisper large-v3 and scored as Character Error Rate (lower is better; ~8–10 is the floor because the scorer writes "9" for "nine").

Trained targets

Language	code
English	en
Twi / Akan	tw
Nigerian & Ghanaian Pidgin	pcm

Also spoken (carried through training, Latin-script):

Language	code	read-back CER
Dutch	nl	7.5
Danish	da	8.9
French	fr	9.1
Malay	ms	13.0
Italian	it	13.3
Swedish	sv	18.9
Spanish	es	19.1
Norwegian	no	19.4
Swahili	sw	20.0
Portuguese	pt	22.2
German	de	28.3
Turkish	tr	29.7

Numbers are for this checkpoint; these languages are intelligible on both Sema Sauti v1 and v2, a few points stronger on one or the other.

Not supported: Arabic, Chinese, Greek, Hebrew, Hindi, Japanese, Korean and Russian — these scripts aren't in the model's tokenizer, so it can't render them. Finnish and Polish come through only partially.

How we built it

• We fine-tuned only the pronunciation stage — the part that carries accent, rhythm and pronunciation — on several hours of Twi and Nigerian Pidgin speech, many speakers, short clips, with Ghanaian Pidgin and African-accented English in the mix.
• The voice encoder and vocoder are left frozen, so the model keeps clean zero-shot voice cloning and clean English while gaining natural Twi and Pidgin.

The shipped weights are step 3300 — the bottom of the evaluation-loss curve, where the voice reads most naturally and the read-back error is lowest, not the last step.

Use

from huggingface_hub import snapshot_download
from chatterbox.tts import ChatterboxTTS
import torchaudio

d = snapshot_download("travis-moore/sema-sauti-v1", token=HF_TOKEN)
model = ChatterboxTTS.from_local(d, device="cuda")
wav = model.generate(
    "Me da wo ase paa. Chale, the meeting go start at nine o'clock.",
    audio_prompt_path="reference.wav",   # 7-10 s of the target voice
    exaggeration=0.5, cfg_weight=0.5,
)
torchaudio.save("out.wav", wav, model.sr)

The folder is a self-contained model directory; it loads with ChatterboxTTS.from_local.

Limitations

• Twi and Nigerian Pidgin are strongest; Ghanaian and Cameroonian Pidgin come through by transfer and are lighter. A matching voice reference helps.
• The "also spoken" languages are ability the model retained, not a dedicated fine-tune — expect a light accent and occasional slips.
• Numbers, acronyms and dates are a known soft spot.
• Clone only voices you have the right to use; outputs carry an inaudible watermark.

Safety

This model clones a voice from a few seconds of audio. Use it only on voices you have the right to use. Impersonating a real person without their consent, deception and fraud are against the intent of this release; disclose synthetic audio where a listener could reasonably be misled.

Versions

Version	Checkpoint	Notes
sema-sauti-v1	step 3300	lowest read-back error; recommended
sema-sauti-v2	step 5124	full training run, final step

---

Fine-tuned from ResembleAI/chatterbox (MIT).