Kakara: Kalenjin text-to-speech

Kakara speaks Kalenjin. It is Kokoro-82M (82 M parameters) fine-tuned on about 290 hours of Kalenjin speech, and it runs faster than real time on an ordinary laptop CPU, with no GPU and no PyTorch.

Three voices: female, male, and blend (an average of many speakers).

Use

pip install onnxruntime numpy soundfile huggingface_hub
from huggingface_hub import snapshot_download
import sys; path = snapshot_download("KalenjinAI/kakara"); sys.path.insert(0, path)
from kakara import Kakara

tts = Kakara(path)
tts.save("Chamgei, amache ityem modelit ni kibare Kakara.", "out.wav", voice="female")   # 24 kHz WAV
audio = tts("Abaibai amun keib kasarta ityem.", voice="male")                       # numpy array

Command line, from the downloaded folder: python kakara.py "Chamgei, amache ityem modelit ni kibare Kakara." --voice male --out out.wav

Text is converted to phonemes by rule (kln_g2p.py), with English loanwords recognised and pronounced as English (kln_english.py). Numbers are read out in Kalenjin. No espeak is needed.

PyTorch users can load kakara.pth with kokoro.KModel(repo_id="hexgrad/Kokoro-82M", config="config.json", model="kakara.pth") and pass torch.from_numpy(np.load("voices/female.npy"))[len(phonemes) - 1] as the voice.

Samples

Generated with speed=1.3.

female male blend
Kongoi missing en toretet ne koigonan amun koanetkee tuguk chechang' chebo muswoknatet.
Ye kiger sugul en arawet ne isibu, atinye tuiyet nebo chaik en Kericho tarigit taman ak sisiit.
Kaitit yu missing, kaatuch gee ak branketisiek mut che nyigisen ago kaitit kora.

Evaluation

20 held-out sentences (never seen in training) from one speaker, compared with the real recordings of those sentences and with our previous model.

System Judge CER ↓ UTMOS ↑ Speaking rate (phonemes/s)
Real recordings (held-out speaker) 0.129 2.81 10.7
Kalenjin TTS v1 (F5-TTS) 0.092 2.69 12.0
Kakara Β· female 0.075 3.25 8.7
Kakara Β· male 0.066 3.25 9.3
Kakara Β· blend 0.070 3.34 8.2
  • Judge CER: character error rate of Rekody/whisper-large-v3-turbo-kalenjin transcribing the audio. Lower means easier to understand. The real recordings score worse than the synthetic voices because they are phone recordings with room noise and natural slurring; synthetic speech is clean and evenly paced.
  • UTMOS (utmos22_strong): a predicted 1–5 listener naturalness rating. It was trained on English and Japanese listening tests, so it reflects audio quality, not how native the Kalenjin sounds.
  • Speed on CPU (laptop, onnxruntime, fp32): real-time factor 0.18 on all cores and 0.56 on one core (seconds of compute per second of audio). TTS v1 (F5-TTS) needs 13–17 s per second of audio on the same CPU.

Limitations

  • Pace. Speech is about 15 % slower than the real speakers, and short phrases (a word or two) sound slowest. Use speed=1.1–1.3 for short phrases; 1.0 suits longer sentences.
  • Pronunciation comes from spelling rules. Names, loanwords spelled the Kalenjin way, and words whose tone or vowel length is not written may be read wrongly.
  • Intonation is learned from read and conversational recordings. Questions and emotion are only partly expressed.
  • Only tested on Kalenjin as spelled in the training transcripts; other spelling conventions may be read differently.

Training

  • Recipe: semidark/kikiri-tts (StyleTTS2 training for Kokoro), adapted to Kalenjin.
  • Stage 1: all 552 speakers (about 290 h), 4 epochs.
  • Stage 2: fine-tuned on five selected voices.
  • Release checkpoint: step 2,882 of stage 2, before the adversarial phase; it scored best on the held-out benchmark.
  • Hardware: one RTX PRO 6000 GPU, about a day of training in total.

Data and credit

Trained on the ANV Kalenjin dataset (Anv-ke/Kalenjin), licensed CC BY 4.0. Thank you to ANV and to everyone who recorded their voice for it. The base model is Kokoro-82M by hexgrad (Apache-2.0).

Responsible use

The voices are derived from real people who recorded speech for a public dataset. Do not use Kakara to impersonate anyone, to deceive listeners, or to put words in a real person's mouth. Label synthetic speech as synthetic where listeners could mistake it for a person.

Downloads last month
10
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for KalenjinAI/kakara

Quantized
(128)
this model

Dataset used to train KalenjinAI/kakara