Ottema

company

https://www.ottema.com.br

AI & ML interests

Brazilian Portuguese NLP Named Entity Recognition Information Extraction Open-Vocabulary NER Structured Generation Agent Reliability Tool Calling JSON Repair Speech Recognition Small Language Models Edge AI Applied Machine Learning

Recent Activity

luis-otte updated a Space 19 days ago

ottema/nemotron-3.5-asr-ptbr-webgpu

luis-otte published a Space 19 days ago

ottema/nemotron-3.5-asr-ptbr-webgpu

luis-otte updated a model 19 days ago

ottema/nemotron-3.5-asr-ptbr-onnx-int4

View all activity

Organization Card

Community About org cards

Ottema

Open AI models for Brazilian Portuguese and reliable production systems.

Ottema develops specialized open-source AI models for Brazilian Portuguese — from streaming speech recognition highlighted by NVIDIA to open-vocabulary information extraction and reliable structured output for AI agents.

Based in Brazil. Built for practical deployment, transparent evaluation, and reproducible research.

🌐 Website · 🤗 Models and datasets · 📦 Open Models Collection

Featured: Nemotron 3.5 ASR for Brazilian Portuguese

ottema/nemotron-3.5-asr-ptbr is Ottema's Brazilian Portuguese adaptation of NVIDIA's 600M-parameter Nemotron 3.5 streaming ASR model.

The model was fine-tuned on approximately 366,000 PT-BR speech samples, representing around 264 hours of audio from CORAA, Multilingual LibriSpeech, and Mozilla Common Voice.

Highlights

🇧🇷 Specialized for Brazilian Portuguese
⚡ Native cache-aware streaming
🎚️ Configurable chunk sizes from 80 ms to 1.12 s
🧠 600M-parameter FastConformer-RNNT architecture
🔧 Compatible with NVIDIA NeMo
🎧 Supports batch and low-latency streaming transcription
🏢 Designed for contact centers, meetings, voice interfaces, and private speech-processing pipelines

🔥 Highlighted by NVIDIA: Ottema was named as the Brazilian Portuguese contributor in NVIDIA's community roundup celebrating the first 42 adaptations built on Nemotron 3.5 ASR.

CORAA evaluation

All models were evaluated on the same 12,676 samples (approximately 17.6 hours) from the official CORAA test split, using the same text-normalization procedure.

Model	Normalized WER ↓	Normalized CER ↓
Ottema Nemotron 3.5 ASR PT-BR	20.96%	10.89%
NVIDIA Nemotron 3.5 ASR zero-shot	31.87%	20.64%
Whisper Large v3	25.15%	14.77%

FLEURS PT-BR evaluation

Model	Normalized WER ↓	Normalized CER ↓
Ottema Nemotron 3.5 ASR PT-BR	10.37%	5.16%
NVIDIA Nemotron 3.5 ASR zero-shot	13.86%	8.31%

Results are self-reported under the evaluation procedure documented in the model card. Differences in normalization and decoding settings may prevent direct comparison with results published elsewhere.

View model, usage, benchmarks, and limitations →

Brazilian Portuguese extraction

Open-vocabulary NER and evidence extraction for formal, informal, and operational Brazilian Portuguese.

Model	Use case	Highlights
`gliner2-ptbr-harem`	NER for formal and journalistic PT-BR	Best entity F1 among the models compared in our HAREM evaluation; approximately 4× faster than the evaluated BERT-CRF baseline
`gliner2-ptbr`	General-purpose NER for chat, customer service, and informal PT-BR	Entity F1 of 0.9976 on our synthetic evaluation set
`gliner2-ptbr-ontoevidence`	Ontology-guided evidence extraction	Hard-negative training designed to reduce unsupported positive predictions

HAREM evaluation

gliner2-ptbr-harem is optimized for Brazilian Portuguese NER using HAREM and related data.

Metric	Result
Entity F1 — per-sample macro	0.4749
Entity F1 — global micro	0.4501
Evaluated latency	31 ms
Evaluated BERT-CRF latency	131 ms

See the model card for the complete evaluation protocol, hardware, baselines, and limitations.

Informal and operational Portuguese

gliner2-ptbr targets conversations, support tickets, customer service, and other less-formal domains.

The model reached an entity F1 of 0.9976 on its synthetic evaluation set. Evaluation on real domain-specific data is recommended before production deployment.

Ontology-guided evidence extraction

gliner2-ptbr-ontoevidence extracts evidence spans using a supplied ontology and includes hard-negative training to improve rejection behavior.

Metric	Result
F1 on the OntoEvidence test set	0.32
Average predictions per text	4.4

Companion dataset:

ottema/gliner2-ptbr-ontoevidence-data contains 2,268 samples across three splits, including multi-label spans and hard-negative examples.

Reliable structured output

Large language models can return malformed JSON, invalid tool calls, or outputs that violate the requested schema.

StructFix explores a compact, specialized recovery layer for these failures.

Resource	Description
`structfix-codet5p-220m`	Repairs malformed JSON and tool-call output against a target schema
`structfix-bench`	250,000 schema-guided examples with controlled noise and constraint coverage
`constraint-dsl`	Compact DSL for declaring typed constraints over JSON outputs

Try the models

GLiNER2 PT-BR Demo

Test open-vocabulary entity extraction with multiple models, label presets, and Brazilian Portuguese examples.

Launch the GLiNER2 demo →

StructFix Demo

Repair malformed JSON and tool-call outputs against typed target schemas.

Launch the StructFix demo →

Lightweight and edge ASR

ottema/stt_pt_quartznet15x5_ctc_small is a lightweight Brazilian Portuguese ASR research baseline based on QuartzNet15x5.

It is intended as a reference for CPU inference, ONNX experimentation, and resource-constrained deployments.

For higher-quality streaming ASR, use ottema/nemotron-3.5-asr-ptbr.

How we work

Production relevance: We evaluate latency, domain-specific accuracy, and robustness—not only aggregate benchmark scores.
Transparent limitations: Model cards document known failure modes, trade-offs, and unsuccessful experiments.
Reproducibility: We publish evaluation details, datasets, and scripts whenever licensing permits.
Responsible data use: We do not train on customer or private data. Published data is synthetic or openly licensed.
Specialization over scale: We build focused models for well-defined operational problems.