kokoro-sv — Swedish voices for Kokoro-82M

Twelve named Swedish voices with dynamic prosody — one line to speak. The model and the neural Swedish G2P download from HuggingFace on first use (cached). No training, no local data, no CUDA required.

Install

pip install kokoro-sv

Use

from kokoro_sv import SwedishKokoro

tts = SwedishKokoro()                       # downloads model + G2P from HF (first run)
tts.speak("Hej, jag är CandyTron!", voice="Stina", out="hej.wav")
print(tts.voices)                           # the 12 voice names

# blend two voices into an in-between one
tts.speak("God morgon!", voice=tts.blend("Björn", "Nils", 0.7), out="mix.wav")

# borrow any Kokoro voice for fun — Swedish words in a French accent 🇫🇷
tts.speak("Bonjour!", voice=tts.kokoro_voice("ff_siwis"), out="fr.wav")

# raw audio (24 kHz float32 numpy) instead of a file
audio = tts.synthesize("Ett, två, tre.", voice="Ebba")

CLI

kokoro-sv voices
kokoro-sv speak "Hej, hur mår du?" --voice Stina --out hej.wav
kokoro-sv blend "God morgon"       --a Björn --b Nils --mix 0.7 --out mix.wav

Voices

Female: Alice · Elsa · Ebba · Stina · Greta · Klara
Male: Björn · Lars · Nils · Anton · Oskar · Viktor

Klara and Viktor each have their own fine-tuned model (livelier prosody), which downloads on first use (~330 MB each). Numbers are read out in Swedish ("CandyTron 4000" → "fyratusen").

Voices interpolate (blend) for intermediate identities and intensity. The male voices are a touch soft on this v1 base (a gender-balanced v2 is planned).

Notes

• No GPU needed — falls back to CPU; Kokoro-82M is tiny, so it's fast.
• First run downloads ~330 MB (model + voices), cached afterward; Klara and Viktor add ~330 MB each the first time they speak.
• Python 3.10+.
• On macOS/x86 pip install kokoro-sv pulls kokoro/misaki cleanly. (On aarch64 Linux, misaki/spacy have no wheels — install with --no-deps and add deps manually.)

More

Training, the reproducible multi-speaker pipeline, evaluation, and the full story: https://github.com/joakimeriksson/kokoro-sv

Voices trained from CC0 data (NST + Swedish LibriVox); Klara and Viktor from Swedish speech synthesized with VoxCPM2. Built on Kokoro-82M (Apache-2.0), StyleTTS2, and kikiri-tts. Apache-2.0.