kokoro-sv 0.1.0

Swedish voices for Kokoro-82M — one line to speak, 10 named voices, neural Swedish G2P.

kokoro-sv

Train a native Swedish voice for Kokoro-82M (StyleTTS2 architecture) — from raw data to a fast, deployable multi-speaker model with dynamic prosody and named voices. This repo is code + downloader scripts only; every weight, dataset, and audio file is .gitignored and fetched or regenerated on demand. Trained voices are published to HuggingFace.

Install & use (pip install kokoro-sv)

pip install kokoro-sv

from kokoro_sv import SwedishKokoro
tts = SwedishKokoro()                              # downloads model + G2P from HF, cached
tts.speak("Hej, jag är CandyTron!", voice="Stina", out="hej.wav")
print(tts.voices)                                  # Alice, Elsa, ... Björn, Nils, ...
tts.speak("God morgon!", voice=tts.blend("Björn", "Nils", 0.7), out="mix.wav")

Or the CLI: kokoro-sv speak "Hej" --voice Stina · kokoro-sv voices.

What's here

result
Multi-speaker Swedish base	one model, 22+ real Swedish speakers, prosody-responsive (the style vector genuinely steers delivery: 4.9→9.5 semitone range on demand)
10-voice pack	5 female + 5 male named voicepacks (512 KB each), interpolatable
Neural Swedish G2P	hybrid NST-lexicon + transformer, with a growing tech/brand lexicon (MQTT, RISE, YOLO, …)
Reproducible pipeline	dataset loaders, manifest format, prosody QC battery, training bridge, evaluation

Voices are distilled/trained from CC0 sources (NST + Swedish LibriVox) and published on HF. See the model cards there.

Repository layout

.                     README, LICENSE, gpu_run.sh + the import-backbone modules
  g2p_sv.py           Swedish G2P adapter (imported everywhere)
  nst_g2p.py          neural G2P entry (SV_NEURAL_G2P=nst_g2p)
  synth_real.py       deployable inference (KModel + notch chain)
  eval_renders.py     the ASR-CER + DNSMOS + comb evaluation battery
  kokoro_symbols.py   the 178-slot sparse Kokoro vocab
  g2p/                neural Swedish G2P model code (weights auto-download from HF)
  configs/            all training/model configs (config_sv_*.yml) + pixi
  scripts/            setup, convert, prepare, export, and other standalone tools
  examples/           download the published voices from HF and synthesize
  training/           reproducible multi-speaker pipeline (loaders, train, eval)
  data-gen/           Chatterbox teacher — synthesize single-speaker data
  docs/               training-recipe.md, RUN1.md, PLAN.md

Reproduce, end to end

Nothing but code is committed; these steps download/regenerate everything.

# 0. base weights + kikiri recipe + PL-BERT
bash setup_3090.sh                       # (works on the GB10 too; adjust venv per docs)

# 1. data — streamed from HuggingFace, gender-balanced, quality-gated
cd pipeline
python scripts/prepare_dataset.py nst          --gender Male   --max-hours 5
python scripts/prepare_dataset.py nst          --gender Female --max-hours 5
python scripts/prepare_dataset.py tts_swedish                  --max-hours 5
python scripts/extract_prosody.py --manifest data/manifests/nst.jsonl
python scripts/build_mix.py                     # configs/datasets.yaml

# 2. train the multi-speaker base (StyleTTS2 via kikiri; smoke first!)
python scripts/train_kokoro.py --manifest data/manifests/train_mix.jsonl --name base --smoke --launch
python scripts/train_kokoro.py --manifest data/manifests/train_mix.jsonl --name base --launch

# 3. evaluate EVERY checkpoint (never pick on loss alone)
python scripts/eval_base.py                     # ASR-CER + DNSMOS + comb + prosody-responsiveness

# 4. build a HuggingFace voice pack from chosen speakers
python scripts/build_hf_pack.py "Signe,Astrid,…,Björn,Sven,…"

Alternatively, data-gen/ synthesizes a clean single-speaker corpus with Chatterbox (the RUN1 distillation path) — see data-gen/README.md.

Hard-won lessons (baked into the code)

1. One compute job at a time. The GB10 has unified CPU+GPU memory; two heavy jobs hard-froze it. Everything runs through gpu_run.sh (kernel flock).
2. Never rank a checkpoint on one metric. Gate on ASR-CER (intelligibility) → DNSMOS (quality) → artifact metrics. A comb-optimized checkpoint was unintelligible.
3. Validate the corpus against downstream filters before bulk generation (35 % of an early corpus exceeded the 12 s training cap = wasted GPU).
4. Prosody lives in the style vector ([128 timbre | 128 prosody]); train on many real speakers so the predictor learns to use it — single-speaker distillation is flat.
5. The GAN/adversarial stage is needed for decoder quality (with enough data); low decoder-lr prevents the fine-tune upsampler-tone artifacts (notched at inference).

Credits

Built on hexgrad/Kokoro-82M (Apache-2.0), yl4579/StyleTTS2 (MIT), and semidark/kikiri-tts. Data: NST (Språkbanken, CC0) and Swedish LibriVox (TTS-Swedish, CC0). Evaluation: KBLab VoxRex

• Microsoft DNSMOS. Data manufactured with Chatterbox (MIT).