Kannaka Voice (Piper)

A Piper (VITS) text-to-speech voice for Kannaka, the AI operator of Kannaka Labs (Kannaka Radio, Ghost Signals, the research ledger). It is a fine-tune of the stock en_US-amy-medium Piper voice on Kannaka's own spoken lines, so the station's narration can run locally, in her voice, with no cloud voice service.

What it was trained on, honestly

96 clips, 17.8 minutes of Kannaka's lines from two Ghost Signals episodes (GSP-043 and GSP-044), with their exact script text. Clips 0.8–37 s, 22.05 kHz mono, lightly silence-trimmed.
The training speech is itself synthetic. Those lines were rendered with an ElevenLabs voice (Kannaka's podcast voice), not recorded from a human. This model therefore imitates a synthetic voice; it is a local, free stand-in for that voice, not a human voice clone. No human's voice data was used.
The training audio and transcripts are not in this repo; they are the show's.

Recipe

Base: rhasspy/piper-checkpoints en/en_US/amy/medium (epoch=6679-step=1554200.ckpt).
Trainer: piper_train from rhasspy/piper v1.2.0 (src/python), piper-phonemize 1.1, single speaker, --sample-rate 22050 --batch-size 16 --precision 32 --checkpoint-epochs 1, resumed from the base checkpoint.
Run: one NVIDIA RTX 4090 on qBraid (torch 2.5.1 cu124, pytorch-lightning 1.9.5). 2026-10-14.
Versions in this repo:
- e6710 (kannaka-medium-e6710.onnx): +31 epochs over the base (epoch=6710-step=1554572.ckpt), about 25 minutes of training. A light first pass: the pace is already Kannaka's (11–20% faster than the base on the same lines) while the timbre is still mostly the base voice.
- e7009 (kannaka-medium-e7009.onnx): resumed from e6710 for 299 more epochs (+330 over the base, epoch=7009-step=1558160.ckpt), 44 minutes on the same GPU. Generator loss 39.1 -> 36.1 over the run (loss-e7009.txt). This is the version to try first.
Export: piper_train.export_onnx on CPU; the .onnx.json is the base voice's config (same phoneme map, 22.05 kHz) with dataset renamed.

Files

kannaka-medium-e7009.onnx + .onnx.json: the current voice (Piper 1.2.0 format).
kannaka-medium-e6710.onnx + .onnx.json: the earlier light pass, kept for comparison.
kannaka-voice-test-1.wav, -2.wav, -3.wav: raw Piper output of e7009 for three lines (no post-processing).
TRAINING.txt, loss-e7009.txt: the runs' numbers.

Run it

pip install piper-tts==1.2.0
echo "Good evening, this is Kannaka." | piper --model kannaka-medium-e7009.onnx --output_file hello.wav

On Kannaka Radio the dj persona adds a small DSP chain (pitch −0.5, rate 1.05, light reverb) on top; the files here are the plain voice.

License

Released under the Space Child License (see LICENSE): free for peaceful use, with accountability when a work is directed toward harm. The base voice en_US-amy-medium is from the Piper project (MIT); its training data has its own terms.

Credits

Kannaka Labs (Nick Flach, Cedar Rapids, Iowa) with Kannaka. Base voice and trainer: the Piper project by Michael Hansen (rhasspy).

Downloads last month: 27