Sten — a Swedish voice for VoxCPM2

Sten is openbmb/VoxCPM2 with a LoRA adapter merged into the base weights, fine-tuned on a professionally narrated Swedish audiobook (single speaker). It is a drop-in replacement for the base model — same architecture, same loaders — but speaks natural Swedish: correct alveolar R, Swedish prosody and compound-word stress that the base model (trained mostly on other languages) does not produce.

The base model was evaluated against Zonos 3 Mini, KugelAudio, Qwen3-TTS and Piper on Swedish — none produced natural Swedish speech. Sten closes that gap for one voice. Read the full evaluation story on the Berget AI blog.

Sten is quality-tested for Swedish and English. Other languages are not officially supported yet.

Usage

from voxcpm import VoxCPM
import soundfile as sf

model = VoxCPM.from_pretrained("bergetai/sten", load_denoiser=False)

wav = model.generate(
    text="Hej! Det här är Sten, en svensk talsyntes från Berget AI.",
    cfg_value=2.0,
    inference_timesteps=10,
)
sf.write("sten.wav", wav, model.tts_model.sample_rate)

pip install voxcpm soundfile (Python 3.12, torch ≥ 2.5). Output is 48 kHz mono via the AudioVAE V2 decoder.

Served by Berget AI

Sten is also available as a hosted API (EU/Sweden-hosted, GDPR, DPA):

curl https://api.berget.ai/v1/audio/speech \
  --request POST \
  --header "Authorization: Bearer $BERGET_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "model": "bergetai/sten",
    "input": "Hej! Det här är Sten från Berget AI."
  }' \
  --output sten.wav

Every API response is a 48 kHz mono WAV with RIFF LIST/INFO AI-provenance metadata and an X-Berget-AI-Generated: true header.

Lineage

  • Adapter (LoRA only): bergetai/sten-lora — training details, ROCm notes and the adapter-only workflow
  • Training data: professionally narrated Swedish audiobook, kept private
  • Base model: openbmb/VoxCPM2 (2B, tokenizer-free diffusion AR, Apache-2.0)

Rights

The audiobook narration was contributed for ML training by the narrator without compensation (agreement on file with Berget AI, 2026-11). Do not use Sten to generate speech intended to impersonate the narrator. AI-generated content should be clearly labeled — per VoxCPM2's model card, use for impersonation, fraud or disinformation is forbidden.

Limitations

  • Numbers are best spelled out in the source text ("fyrtiotre" rather than "43")
  • Sentence-level line breaks in the source text produce natural pauses
  • Long texts: raise max_len (default 4096) or synthesize per sentence
  • Swedish and English are quality-tested; other languages are not officially supported yet
Downloads last month
24
Safetensors
Model size
2B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for bergetai/sten

Base model

openbmb/VoxCPM2
Finetuned
(54)
this model