Writing
From the workshop
Notes on voice AI, OpenVoiceOS development, synthetic data, and minority-language speech technology.
- 7 min read
Exporting and Quantizing Open Speech Models So They Actually Run
A trained speech model on a research GitHub page is not a voice assistant. We convert open ASR and TTS checkpoints to ONNX, CoreML and GGUF, quantize them, and validate the output. Then we publish the results under OpenVoiceOS so every language they cover ships in a real, offline assistant.
- ONNX
- CoreML
- GGUF
- ASR
- 12 min read
Porting With Machines, and the Licence Question We Could Not Answer
We rewrote several C, C++ and Java programs (espeak-ng's G2P, Cotovia, AhoTTS, HermiT) as pure Python, with an AI reading the original source and a human orchestrating. No human on our side read the originals. That raises two separate questions: can the output be owned at all, and is it a derivative of the input? We kept the upstream licences because that was cheaper than answering. We still think the question is open.
- FOSS
- Licensing
- Open Source
- Python
- 8 min read
A Family of Pure-ONNX Speech Libraries
TigreGótico maintains a set of speech libraries — bandwidth extension, voice cloning, speaker embeddings, VAD, word stress, phonemization, TTS, and a metrics library to score them all — that share one runtime rule: onnxruntime and numpy only, no PyTorch, no GPU required.
- ONNX
- TTS
- voice cloning
- VAD
- 10 min read
How the Phonology Stack Fits Together
An architecture tour of our text-to-pronunciation stack: scriptconv for notation, orthography2ipa as the cross-language grapheme-to-IPA engine, language-specific frontends built on top of it for Portuguese, Basque, Mirandese, Barranquenho and Arabic, and phonematcher for sound-based search. Shows why the layers exist, what a candidate lattice is, and real dialect output.
- G2P
- IPA
- Phonetics
- NLP
- 9 min read
Choosing a Voice-Cloning Engine
voiceclonnx runs 10 voice-conversion engines behind one API, from kNN feature-swap to AR codec-LM. This is a guide to the model families behind them, the real measured trade-off between intelligibility and speaker similarity, and how to pick an engine for a specific job.
- ONNX
- voice cloning
- voice conversion
- self-hosted
- 10 min read
Cleaning Up Bad Audio: Denoising and Bandwidth Extension in audiosronnx
audiosronnx treats denoising and bandwidth extension as two separate jobs: the real engine registry, model sizes and licenses, the rejected-models list, and how to check whether the output actually improved.
- ONNX
- denoising
- bandwidth extension
- speech
- 9 min read
Measuring Speech Quality Without a Listening Panel
A working guide to speechonnxmetrics: what MOS, no-reference MOS predictors, intrusive signal metrics and ASR-based WER/CER actually measure, when each applies, real scores from real audio, and why a predicted MOS is evidence, not truth.
- speechonnxmetrics
- TTS
- ONNX
- evaluation
- 7 min read
Scripts and Phonetic Notations: What scriptconv Actually Converts
A tour of scriptconv, the zero-dependency library that detects writing systems and converts between phonetic notations. Covers IPA, ARPABET, and X-SAMPA, ISO-15924 script detection, Buckwalter transliteration for Arabic, Hangul decomposition into jamo, and kana conversion, with real, executed examples and honest limits.
- IPA
- Phonetics
- NLP
- Linguistics
- 11 min read
Your sentiment model can't tell a complaint from a goodbye
Two angry-looking support messages. One customer is about to escalate; the other is about to leave without a word. Almost no emotion model can tell them apart — because they're all missing the same axis. Introducing emotion-algebra.
- Affective Computing
- Emotion
- Machine Learning
- LILACS
- 4 min read
Why We Hoard Data: From Scraped Catalogues to Smarter Speech and Language Models
Clean, typed, well-provenanced data is the raw material of every model we ship. How the catalogues our scrapers build become ASR biasing vocabularies, intent classifiers, synthetic NER corpora, G2P lexicons, TTS voices, and honest fuel for LLMs.
- Datasets
- Data Collection
- ASR
- NLP
- 6 min read
If Everyone Ships an App, We'll Ship Voice: Turning Websites Into Voice Apps
Every site that matters got wrapped in a mobile app. We propose the opposite move for the voice and CLI era: a clean API plus a voice skill per site, so the web becomes browsable by ear and by keyboard. One site at a time, it adds up to a voice browser.
- Voice
- Accessibility
- OpenVoiceOS
- Web Automation
- 2 min read
Usenet in 2026: a clean, pre-AI text corpus for training and evaluation
Usenet is a pristine archive of pre-AI human discourse: decades of newsgroup posts, all human-written, none of it touched by language models. That makes it valuable training and evaluation data for language and speech models. We built a small Python tool to harvest it.
- Usenet
- Datasets
- NLP
- 3 min read
Two Voices, Every Language: Miro & Dii
TigreGótico is partnering with OpenVoiceOS to give the assistant two consistent voice identities, Miro and Dii, that sound the same in every language, built with our voice-cloning technology and the phoonnx engine. Two TTS models for every language someone requests, endangered tongues included.
- phoonnx
- TTS
- OVOS
- voice cloning
- 4 min read
Saying It Right: Disambiguating Portuguese Heterophones for TTS
Many European Portuguese words are spelled the same but pronounced differently depending on meaning, and getting the vowel wrong makes a TTS voice say the wrong word. We built bifonia-pt-homographs, an open meaning-labelled dataset of 56,891 sentences over 27 words, and a tiny zero-dependency resolver that hits ≈94% where heavyweight POS taggers plateau at ≈75%.
- Datasets
- Portuguese
- TTS
- Grapheme-to-Phoneme
- 5 min read
TTS Models That Run on a Potato
phoonnx is a research framework for VITS-based text-to-speech built to run comfortably on low-end hardware. No GPU, no cloud, no API key: just a ~15.65 million-parameter ONNX voice and a CPU. Here is how small a good voice can be, and how we train them.
- phoonnx
- TTS
- ONNX
- VITS
- 5 min read
Introducing Our Music Database Scrapers
A tour of the family of typed Python clients we maintain for music sources — Bandcamp, SoundCloud, SomaFM, TuneIn, iHeartRadio, and the great music encyclopedias — all emitting consistent, typed media metadata behind one clean interface and riding the same compliant, low-volume HTTP transport.
- Scrapers
- Media Metadata
- Music
- Python
- 2 min read
A Multilingual Sentence-Types Dataset: Questions, Commands, Statements
We published sentence-types-multilingual — nearly 70,000 sentences across seven languages, classified by grammatical type (question, command, statement, exclamation). It is the training corpus behind the little_questions routing library.
- Datasets
- Multilingual
- NLP
- Intent
- 6 min read
Composable, Drop-In requests Sessions for Resilient Public-Data Access
Two composable requests.Session subclasses for reading public web pages reliably without a headless browser in the hot path: TLS-compatible transport, a FlareSolverr proxy for JS challenges, a Wayback Machine fallback, and IP-diversified requests: unblock_requests and anon_requests.
- HTTP
- Scraping
- Cloudflare
- Anti-Bot
- 3 min read
Robots.txt, Sitemaps, and Ethical Web Scraping
Before you build a scraper, scout the site. sitemapper reads robots.txt, fetches every sitemap, and optionally crawls the link graph, so your scraper starts from the site's own contract instead of brute force.
- Web Scraping
- Sitemaps
- Ethics
- Robots.txt
- 5 min read
Classical NLP for Portuguese: Syllabification and Grapheme-to-Phoneme
A look at our rule-based, fully offline Portuguese NLP stack: silabificador for syllabification and TugaPhone for dialect-aware grapheme-to-phoneme, and how they connect to the broader orthography2ipa work for Lusophone varieties. No deep-learning black boxes: deterministic, fast, and dependency-light.
- NLP
- Portuguese
- Phonemization
- Grapheme-to-Phoneme
- 8 min read
Grapheme-to-IPA for 820 Languages
orthography2ipa is a pure-data, linguistically grounded resource that maps spelling to IPA and models how phonemes surface as allophones across 909 language specs, 820 languages, and 20+ language families. A candidate lattice, a maximal-munch tokenizer, phonological and script distance metrics, dialect lineage, and a schema-validated spec set cited to the dialectological literature — no trained weights, fully self-hostable.
- G2P
- IPA
- Phonetics
- NLP
- 2 min read
Introducing the First Phonemizer for Barranquenho
g2p_barranquenho is the first open grapheme-to-phoneme converter for Barranquenho, the Ibero-Romance contact language of Barrancos, Portugal — rules derived from the municipality's newly published orthographic convention, auditable against the committed sources.
- Phonemization
- Barranquenho
- Minority Languages
- NLP
- 3 min read
Cloning Voices for Endangered Languages: Building a Text-to-Speech Model for Asturian and Aragonese
We are releasing experimental text-to-speech models for Asturian (ast) and Aragonese (an), two minority Romance languages with essentially no commercial voice coverage. We built them with a hybrid pipeline: filtered Common Voice data, zero-shot revoicing, and VITS training via phoonnx.
- TTS
- Asturian
- Aragonese
- Minority Languages
- 3 min read
OVOS & HiveMind in the Manufacturing Industry
The COALA and WASABI EU projects have built an entire industrial voice-assistant framework around OVOS + HiveMind, integrating them with their own tools, UI, and conversation engines.
- OVOS
- HiveMind
- Industry
- manufacturing
- 1 min read
Synthetic Wakeword Datasets: Seven Assistant Names, One Detector
We published seven synthetic wakeword datasets for common voice assistant names: hey_computer, hey_mycroft, hey_siri, alexa, home_assistant, voice_assistant, wake_up. Train a detector that works everywhere.
- Datasets
- Wakewords
- Speech
- Synthetic
- 3 min read
Introducing phoonnx: OpenVoiceOS's New TTS Framework
phoonnx is now the primary text-to-speech framework for OpenVoiceOS: a full VITS/ONNX training and inference stack, starting with new Basque voices for Miro and Dii.
- phoonnx
- TTS
- OVOS
- ONNX
- 4 min read
OpenVoiceOS and Home Assistant: A Voice Automation Dream Team
Home Assistant handles automation. OVOS handles voice. Three integration layers make the combination work: Wyoming bridges for HA's voice pipeline, ovos-persona-server as a conversational agent, and HiveMind for surfacing OVOS devices as native HA entities.
- OVOS
- Home Assistant
- Smart Home
- Voice Automation
- 2 min read
Making Synthetic Voices From Scratch
Creating a voice for a text-to-speech system usually requires a real person to spend hours recording audio. That’s expensive, time-consuming, and in many languages or accents, the voices just don’t exist at all
- TTS
- Synthetic Data
- Voice Cloning
- OVOS
- 2 min read
Miro & Dii TTS Training Data: 40 Open Datasets Across 20 Locales
We published 40 synthetic training datasets for the Miro and Dii voices across Portuguese, Dutch, German, French, Italian, Japanese, Spanish, and more. Every language gets two consistent voice identities, built with voice cloning.
- TTS
- Voice
- Datasets
- Miro & Dii
- 2 min read
No Language Left Behind
Eliminating language barriers in OpenVoiceOS through language detection, translation plugins, and bidirectional translation capabilities.
- OVOS
- multilingual
- language-detection
- translation