Kokoro Swedish — multi-speaker voice pack (12 named voices)

A Swedish Kokoro-82M model fine-tuned on 23 real Swedish speakers (NST + LibriVox, ~8.6 h), shipped with 10 named voicepacks. One model, many voices — each voice is a 512 KB style tensor; the model reproduces each speaker's prosody, so the voices are genuinely dynamic (not the flat prosody typical of single-speaker distillation).

Two more voices, Klara and Viktor, ship with their own fine-tuned weights (see below).

Voices

Klara and Viktor (own models)

A voicepack only sets timbre; the prosody comes from the shared predictor. Klara and Viktor were therefore made differently: each started from a reference take generated with VoxCPM2 voice design, VoxCPM2 cloned that take on ~2,300 Swedish sentences (~2.8 h each, every clip gated on speaker similarity and ASR CER), and the multi-speaker model was fine-tuned on that set (stage 2, 5 epochs) into a separate model per voice. They are livelier than a voicepack on the shared model (F0 variation 7.7 / 7.1 semitones vs 6.5 / 5.6 for their own packs on kokoro_sv.pth), at the same intelligibility.

voices_manifest.json names their weights ("model": "models/<Name>.pth"); load that file instead of kokoro_sv.pth for these two. A Klara or Viktor pack used on the shared model gives their timbre with the flatter shared prosody.

Prosody-responsiveness is measured: feeding a monotone speaker's voicepack yields ~4.9 semitone F0-variation; an expressive one yields ~9.5 — the style vector genuinely steers delivery.

Files

The neural Swedish G2P (required — inference must phonemize identically to training) lives in Joakim/swedish-kokoro.

Usage

import torch, soundfile as sf
from kokoro import KModel
from scipy.signal import iirnotch, filtfilt
from g2p_sv import SwedishG2P            # SV_NEURAL_G2P=nst_g2p

model = KModel(repo_id="hexgrad/Kokoro-82M", config="config.json", model="kokoro_sv.pth").eval()
voice = torch.load("voices/Stina.pt", weights_only=True)   # pick any voice
# Klara / Viktor: model="models/Klara.pth" with voices/Klara.pt
g2p = SwedishG2P(backend="neural")

ipa = g2p("Hej, jag är CandyTron, din godisrobot!")
ids = [i for i in (model.vocab.get(p) for p in ipa) if i is not None]
audio = model.forward_with_tokens(torch.LongTensor([[0]+ids+[0]]), voice[len(ids)-1], speed=1.0)[0].squeeze().numpy()
for f0 in (2400, 4800, 7200, 9600):          # notch the fine-tune upsampler tones
    b, a = iirnotch(f0, Q=35, fs=24000); audio = filtfilt(b, a, audio)
sf.write("out.wav", audio, 24000)

Voicepacks interpolate: 0.6*voiceA + 0.4*voiceB gives an intermediate voice/intensity.

Known limitations (this is v1)

• Male voices are a touch soft. The training mix is ~66% female, so the decoder leans female and renders male voices less fully male than a dedicated model would. A gender-balanced v2 base is planned. Björn is the deepest/most-male here.
• A residual decoder "comb" (~1.5 on our metric vs ~0.4 for single-speaker) is reduced by the four notch filters in the usage snippet; some remains.

Provenance

• Speakers: NST Swedish ASR corpus (CC0, via HF KTH/nst)
  • Swedish LibriVox (datadriven-company/TTS-Swedish, CC0).
• Klara, Viktor: Swedish speech synthesized with VoxCPM2 from one designed reference take each (no real person's voice).
• Training: kikiri-tts two-stage recipe (StyleTTS2), multispeaker, mel-GAN + WavLM SLM, max_len 800, decoder-lr 1e-5. Checkpoints selected by ASR-gated evaluation (KBLab VoxRex CER + DNSMOS), never loss alone.
• Trained on an NVIDIA GB10.

Credits

hexgrad/Kokoro-82M · openbmb/VoxCPM2 · semidark/kikiri-tts · yl4579/StyleTTS2 · Språkbanken/NST · KBLab VoxRex (eval)