Content is what remains
A speech tokenizer that learns from many voices saying the same words. It keeps what they share and lets the speaker, the room and the microphone fall away.
- 1.2%
- speaker probe accuracy on PINT tokens, down from 93.1% on HuBERT
- 27–30%
- lower language-model perplexity than HuBERT and WavLM tokens, same model and data
- 56 b/s
- after dedup and BPE, approaching character-level text BPE at 48
Four voices. The same tokens.
Each row is one person saying "Please call Stella." A colour is a token, so matching colours mean matching tokens. Play a voice, hover a band, then switch to HuBERT and watch the rows fall apart.
14% of runs differ
Averaged over the six speaker pairs (normalized edit distance). Lower means the tokens carry less of the speaker.
- PINT14%
- HuBERT40%
- WavLM47%
Real k=200 output from the PINT repository. Four VCTK speakers not seen in training read "Please call Stella." (VCTK corpus, University of Edinburgh, CC BY 4.0; converted to 16 kHz mono). HuBERT and WavLM have their own 200-token codebooks, so compare how the rows agree, not the colours between tokenizers. "Lined up" leaves a gap where a speaker has a run the others lack.
Results
Tokens that stop carrying the speaker are easier to model. PINT also cuts ABX error by 42% against the baselines.
Use it
16 kHz audio in, one 768-dimensional vector or one token id per 20 ms out. Use the tokens as semantic targets for audio codecs or as input for speech language models. It works on English speech.
Python
# pip install git+https://github.com/nyrahealth/PINTfrom pint_infer import PINTTokenizer, format_runstok = PINTTokenizer.from_pretrained("nyralabs/PINT")ids = tok.encode("p230.wav") # one token id per 20 msframes = tok.embed("p230.wav") # (T, 768) float32print(format_runs(ids))3×35 85×2 68×3 60×4 94×3 96×6 110×1 39×2 132×2 63×2 48×9 173×5 79×2 4×1 98×4 99×1 9×1 175×7 0×4 157×4 34×3 46×4 3×5Research and non-commercial
Weights are free under CC BY-NC 4.0, and the inference code is MIT.
In a product
The released weights are non-commercial. For a product, talk to us about a license.
For researchers
The paper, how to cite it, and where to find the weights and code. PINT is the first published piece of the representation line of our research program.
Content is What Remains: Invariant Speech Tokenization from Parallel Utterances
Laurin Wagner, Bernhard Thallinger, Miroslav Stankovic, Mario Zusag
@misc{wagner2026contentremainsinvariantspeech,
title = {Content is What Remains: Invariant Speech Tokenization from Parallel Utterances},
author = {Laurin Wagner and Bernhard Thallinger and Miroslav Stankovic and Mario Zusag},
year = {2026},
eprint = {2607.19033},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2607.19033},
note = {Accepted at Interspeech 2026}
}- Read the paperarXiv 2607.19033
- The write-upWhy content is what remains, and where it fits
- nyralabs/PINTWeights and six codebooks on Hugging Face
- nyrahealth/PINTInference code and example clips on GitHub
Work with the lab
Research collaborations, data partnerships and shared evaluation. Tell us what you are working on.
Get in touchRelated research
CrisperWhisper 2.0
Verbatim speech recognition for how people actually talk: controllable, multilingual, and precisely timed to the word.
Closing the verbatim data gap
How to upgrade existing spontaneous-speech corpora into faithful verbatim datasets while preserving the transcripts you already trust.
How precisely do speech systems time words?
How precisely do speech systems place word boundaries? An open benchmark that scores 22 forced aligners and ASR systems against 100k hand-labelled words on conversational, stuttered and noisy speech, and measures how fast each one runs.