Ghanaian English TTS (VITS)

Author: Prince Nasamu Alhassan

Overview

Fine-tuned from facebook/mms-tts-eng on ghana_english_single: 549 clips, 70.9 minutes, one speaker.

One speaker on purpose. The larger ghana_english_tts has 25 hours and no speaker labels, and VITS has no speaker conditioning, so training on it averages every broadcaster into one muddy persona. The cost is that 549 rows is a small corpus and the speaker reads formally.

6,000 steps, and that number is not modesty. On 549 rows, 20,000 steps is 583 epochs. The first attempt was queued at 20,000 before anyone checked what that meant for a corpus this size, and was restarted.

It is the weakest link in the English loop, measured. The same sentence was synthesised five times and transcribed by two recognisers. VITS samples noise in its duration predictor, so every take differs:

The text was What is my battery level? every time. Our recogniser recovered it exactly twice in five and the baseline never did — so the recogniser is fine and the voice is not. It mispronounces level consistently enough that both models hear evil.

More steps will not fix this: 6,000 is already 175 epochs over 549 clips. It needs more voice data, and the obvious source — ghana_english_tts, 6,614 rows over 24.8 hours — has no speaker labels, and VITS has no speaker conditioning, so training on it averages every broadcaster into one muddy persona. That is a real dead end, not a to-do.

Use it

import torch, soundfile as sf
from transformers import AutoTokenizer, VitsModel

repo  = "PrinceAlhassanNasamu/tekyerema-tts-eng"
tok   = AutoTokenizer.from_pretrained(repo)
model = VitsModel.from_pretrained(repo).eval()

text = "your text here".lower()          # LOWERCASE IS NOT OPTIONAL -- see below
with torch.no_grad():
    wav = model(**tok(text, return_tensors="pt")).waveform[0]
sf.write("out.wav", wav.numpy(), model.config.sampling_rate)

Read this before sending it real text. These VITS tokenisers are CHARACTER-level, and a character outside the vocabulary is dropped silently -- no error, no gap in the audio.

Measured 2026-09-23 on 6,945 real agent commands (tekyerema-distill-agent-v2, the utterance-disjoint corpus), per voice, not borrowed from a sibling:

voice	raw text	after .lower()	after .lower() + number verbalisation
tekyerema-tts-eng	6.8%	1.2%	0.00%
tekyerema-tts-twi	10.4%	5.5%	0.01%
tekyerema-tts-hau	12.0%	6.8%	0.00%
tekyerema-tts-kus	12.6%	8.0%	1.72%
tekyerema-tts-ewe	12.8%	7.6%	0.00%

Between 98% and 100% of real commands lose at least one character if you send them raw.

The single largest cause is capitalisation, and it was missing from every earlier version of this card. None of these five vocabularies contains A-Z. Uppercase accounts for 36-82% of everything dropped -- 81.8% for English. "Send Kwame 50 cedis" loses its S, its K and its digits. text.lower() is a one-line fix and it is the highest-value thing on this page.

Digits are the second cause. Vocabulary coverage is erratic rather than absent: Ewe and Kusaal have no digits, Hausa has only 6, Twi has only 2 and 3, and English has 0-6 but not 7, 8 or 9 -- which is worse than having none, because it fails invisibly. Spell numbers into words before synthesis.

Kusaal keeps a 1.72% residual after both fixes that the other voices do not: its 35-character vocabulary has no c, so cedis is spoken as edis -- the word naming the money in a payment confirmation. Route words whose letters this voice lacks to a voice that has them. AGENT/agent_core.py does the number and routing halves (verbalize_numbers, sayable_segments); the lowercasing is on you.

Reproduce: overnight/exp003_tts_coverage.py and overnight/exp003b_tts_fix.py in the Tɛkyerɛma repo.

Training data

Trained on the Ghana Speech dataset and related Ghanaian corpora, licensed CC BY-NC 4.0.

Intended use & license

Non-commercial use only (CC BY-NC 4.0). This is inherited from the training data and required by the terms under which the compute was granted: models trained in that window are non-commercial by condition of access, not by inference.

Limitations, stated plainly

• Dagbani did get a recogniser, and the claim that it could not was wrong twice over. Every card on this account used to say that "one fine-tuning session on 74 validation rows would not change that". Those 74 rows are the eng-dag machine-translation validation split; the Dagbani speech data in this same account is waxal_dag — 13,228 training rows, 1,750 validation rows, ~71 hours, 1,041 speakers with the largest at 1%. Trained on it, tekyerema-asr-mms-dag scores 36.94 / 11.71, against the 86.59 / 33.95 this project had believed was the ceiling. It still loses to FarmerlineML/w2v-bert-2.0_2026_dagbani_ASR at 29.20 / 9.27, which is what the agent actually serves. A number carried across from a translation table into a speech claim was then repeated on every card here until 2026-09-22.
• Evaluation is on read and machine-translated text. No recordings of people speaking agent commands in these languages exist. Numbers measured this way are optimistic about phrasing and pessimistic about code-switching, and should not be read as field performance.
• Research work from a hackathon entry, not a supported product.

The rest of the family

Recognisers

• whisper-large-v3-turbo-tekyerema-eng-foundation — Ghanaian English ASR — course 1 (foundation)
• kusaal-whisper-small-lora — Kusaal ASR (Whisper-small LoRA, superseded)
• kasa42-asr — KASA-42 (Kusaal, third-party export)
• tekyerema-asr-ctc — Twi ASR (w2v-BERT CTC)
• tekyerema-asr-mms-ewe — Ewe ASR (MMS adapter)
• tekyerema-asr-mms-dag — Dagbani ASR (MMS adapter)
• tekyerema-asr-mms-hau — Hausa ASR (MMS adapter)
• tekyerema-asr-mms-kus — Kusaal ASR (MMS adapter)
• whisper-large-v3-turbo-tekyerema-eng — Ghanaian English ASR (Whisper large-v3-turbo)

Voices

• tekyerema-tts-twi — Twi TTS (VITS)
• tekyerema-tts-kus — Kusaal TTS (VITS)
• tekyerema-tts-ewe — Ewe TTS (VITS)
• tekyerema-tts-hau — Hausa TTS (VITS)
• tekyerema-tts-eng — Ghanaian English TTS (VITS)

Agent models

• tekyerema-1-reply — Tɛkyerɛma-1 reply adapter (arm ①)
• tekyerema-1-native-reply — Tɛkyerɛma-1 reply adapter (arm ②)
• tekyerema-1-tool — Tɛkyerɛma-1 tool adapter (arm 1)
• tekyerema-audio-native — Tɛkyerɛma-1 audio-native (arm 3)
• tekyerema-audio-native-4k — Tɛkyerɛma-1 audio-native, 4,000 clips (arm 3 v2)
• tekyerema-1-native-tool — Tɛkyerɛma-1 tool adapter (arm 2)

Translation

• tekyerema-nllb600m-v1 — Tɛkyerɛma MT v1 (NLLB-600M)
• kusaal-nllb-600M — Kusaal MT specialist (NLLB-600M)

Routing

• tekyerema-intent-afroxlmr — Intent classifier (AfroXLMR)

Acknowledgements

Compute resources provided by AI Skills and Compute Africa (AISCA). Trained on the Ghana NLP H200 GPU. Please keep derivatives non-commercial and share improvements back with the Ghana NLP community (ghananlpcommunity).