Ottema

AI that runs where your data is.

Speech, reliable agents, and edge inference — built in Brazil.

🎙️ Speech Recognition · 🔊 Speech Synthesis · ⚡ WebGPU & Edge · 🧩 Reliable Agents · 🇧🇷 Brazilian Portuguese

Website · Models · Spaces · Collections

Local Voice AI Stack

Ottema builds browser-native speech systems that keep voice data on the user's device.

Audio → Nemotron ASR → LLM / Agent → Magpie TTS → Audio
        local speech      reasoning       local voice

Both projects run without a Python or inference backend. Model assets are downloaded by the browser and can be reused from its cache.

Browse the complete Ottema Local Voice AI Stack →

What we build

1. Voice AI & Edge

Local-first ASR and TTS for private assistants, contact centers, real-time interfaces, and resource-constrained deployments.

2. Reliable AI Agents

StructFix is a compact recovery layer for malformed JSON and invalid tool-call output. It is supported by StructFix Bench, with 250,000 schema-guided examples, and a live demo.

3. Brazilian Portuguese AI

Two specialized GLiNER2 models for open-vocabulary NER in Brazilian Portuguese:

• GLiNER2 PT-BR HAREM — NER for formal and journalistic text
• GLiNER2 PT-BR — general-purpose NER for conversations and customer service

Our experimental OntoEvidence model and companion dataset explore ontology-guided evidence extraction with hard negatives.

Flagship projects

1. Nemotron 3.5 ASR PT-BR — WebGPU
2. MagpieTTS Web — PT-BR
3. StructFix
4. GLiNER2 PT-BR HAREM

How we work

• Local by design: prioritize on-device and edge inference when it improves privacy, latency, and deployability.
• Production relevance: evaluate domain accuracy, latency, robustness, and operational constraints.
• Transparent limitations: document known failure modes and trade-offs in each model card.
• Reproducibility: publish evaluation details, datasets, and scripts whenever licensing permits.
• Responsible data use: do not train on customer or private data; published data is synthetic or openly licensed.

About Ottema

Ottema is an applied AI company based in Brazil. We build specialized systems that operate close to the data — in the browser, at the edge, and inside reliable production workflows.

Our work builds on open research and technology from NVIDIA, Hugging Face, ONNX Runtime, NeMo, GLiNER, CodeT5+, CORAA, Mozilla Common Voice, Multilingual LibriSpeech, and Linguateca HAREM. See each repository for complete attribution, evaluation methodology, limitations, and license terms.

Licensing is defined individually for each repository.

Hugging Face · Website · LinkedIn