nb_NO-vox-medium

nb_NO-vox-medium is a Norwegian bokmål text-to-speech voice for Piper. It is a distillation of the OpenBMB VoxCPM2 teacher into a small VITS model fine-tuned from the Norwegian talesyntese base checkpoint.

Highlights

• Norwegian bokmål — trained and verified end to end for one language
• Distilled from VoxCPM2 — teacher prosody in a 63.5 MB student
• Real time on CPU — runs on a Raspberry Pi, no GPU needed
• Home Assistant ready — a drop-in Piper voice
• Trained on spoken dates and names — years, dates, times, and common Norwegian names are part of the training data
• Open data lineage — only modern, openly licensed training text

Quick Start

Install

pip install piper-tts

This release was tested with piper-tts 1.8.0.

Piper CLI

piper \
  --model nb_NO-vox-medium.onnx \
  --config nb_NO-vox-medium.onnx.json \
  --output_file out.wav \
  -- "Værmelding for Vestlandet: Regn og frisk bris fra sjøen hele dagen."

Python

import wave
from piper import PiperVoice

voice = PiperVoice.load("nb_NO-vox-medium.onnx")
with wave.open("out.wav", "wb") as wf:
    voice.synthesize_wav("Hei! Dette er en norsk stemme laget lokalt.", wf)

Home Assistant

Copy the ONNX file and the JSON config into the Piper voice directory, then select the voice.

Model Details

Property	Value
Language	Norwegian bokmål (nb, espeak voice nb)
Architecture	VITS (Piper), single speaker
Quality	medium, 22.05 kHz mono
Model size	63.5 MB ONNX
Base model	NbAiLab/nb-tts-norwegian-piper-trimmed (login required) Mirror (Piper talesyntese, epoch 3459, CC0 data; exact base checkpoint: epoch=3459-step=2052250.ckpt)
Training data	10 333 synthetic clips, 13.9 h (10 336 passed verification; 3 dropped)
Fine-tuning	from epoch 3459 to 3759; released checkpoint 3744 (+285), batch 20, LR 1e-4, precision 32

Evaluation

Word error rate on 100 held-out Norwegian Wikipedia sentences, judged with NbAiLab/nb-whisper-large and normalized before comparison:

Model	WER mean	WER median	Perfect clips
VoxCPM2 teacher (2B, GPU, 48 kHz)	11.2 %	7.4 %	40 / 100
This model	12.5 %	9.1 %	31 / 100

The released checkpoint is epoch 3744. Four checkpoints were exported and scored on the same 100 sentences: epoch 3494 14.5 %, 3554 13.2 %, 3629 13.7 %, and 3744 12.5 % mean WER. The final choice used WER, not validation loss.

The remaining gap to the teacher is dominated by the espeak-ng nb frontend and rare foreign names.

Targeted tests, 16 sentences each, same judge:

Test set	This model
Dates and years	10.1 %
Norwegian names	6.0 %

None of the 16 date or name sentences occur in the training text. The 100-sentence main set has 3 overlapping lines.

How it was made

1. Corpus. 12 097 modern sentences: Wikipedia, NDLA, Alpaca sets, Språkrådet Q&A, and drill lists for names, cities, dates, times, and years. Lines from public-domain books were generated first but skipped, because the language sounded archaic.
2. Generation. VoxCPM2 read every sentence with a fixed synthetic speaker ("vox8").
3. Verification. NB-Whisper transcribed every clip. Text comparison used espeak-ng phonemes, so punctuation and number format do not cause false rejects. 10 336 clips passed (85 %); 1 761 were rejected after two retries. Three pathological clips were dropped before training, leaving 10 333.
4. Preprocessing. Each clip was trimmed, normalized to -23 LUFS with a -0.95 dBFS peak ceiling, resampled to 22 050 Hz, and aligned to the base phoneme map.
5. Fine-tuning. Checkpoints were ranked by validation loss, then the final checkpoint was selected by WER on the held-out corpus, not by validation loss.

Provenance and licensing

The audio and the transcripts are not distributed. Only the model weights are published. We found no evidence that training sentences can be recovered from the weights.

Source	Clips	License (as recorded in the build metadata)
Norwegian Wikipedia dump, two Språkbanken sets	5 195	CC BY-SA
NDLA learning paragraphs	1 945	CC BY-SA 4.0
NbAiLab/norwegian-alpaca instruction sets	1 457	CC BY 4.0
SSB name statistics and Simplemaps Norway cities drill lists	1 344	factual data
hhval/norwegian_qa (Språkrådet språkspørsmål og svar)	384	Apache 2.0
Unmatched text	8	not recorded

Component	Origin	License
Base model	NbAiLab/nb-tts-norwegian-piper-trimmed, exact checkpoint epoch=3459, NST Talesyntese data	CC0 data
Teacher model	OpenBMB VoxCPM2	Apache 2.0
Training audio	Synthetic VoxCPM2 output	—
Piper training code	rhasspy/piper	MIT
Piper inference	piper1-gpl	GPL-3.0-or-later
These weights	this fine-tune	CC0 1.0

The weights are CC0 1.0, the same licence as the base checkpoint's data. Use the voice for any purpose, including commercial use. The training text keeps its own licences; it is not distributed with this model. The voice is fully synthetic and is not derived from an identifiable real person.

Samples

Text	Audio
Weather report	sample_1_weather.wav
Date: 17 May 1814	sample_2_date.wav
Names: Jan Hansen	sample_3_names.wav
Wikipedia: Folkteatern	sample_4_wikipedia.wav

Teacher comparison

The full text of the year description is in months.txt. Both voices read the same text:

Voice	Audio
This model	sample_5_months.wav
VoxCPM2 teacher	sample_6_months_teacher.wav

Listen to both files back to back to judge the distillation gap.

Limitations

• Bokmål only. Nynorsk is untested and likely weak.
• Medium quality: good for assistants and read-aloud, not studio audio.
• Prosody is flatter than the VoxCPM2 teacher.
• Rare English loanwords and long numbers can sound wrong.
• Expand numeric dates and times before synthesis when the format is unusual.
• Do not use the voice for impersonation or disinformation. This is an AI-generated synthetic voice.

Credits

• OpenBMB VoxCPM2 — teacher model
• NbAiLab nb-whisper-large — clip verification
• rhasspy/piper — training code
• piper1-gpl — inference
• Språkbanken / Nasjonalbiblioteket — CC0 Norwegian speech
• NDLA, Wikipedia, SSB, Språkrådet — text sources