Building full-duplex Voice AI for European Portuguese.
Tender Touch, Lda. has established a new R&D direction focused on real-time conversational voice AI. Our first programme is European Portuguese — with a longer-term objective of developing a repeatable methodology for other under-served European languages.
An existing Portuguese company. A new R&D direction.
Tender Touch, Lda. is an operating Portuguese company with existing digital activities, including AUTO.MOTO.pt. In 2026, the company began establishing a dedicated AI & Voice R&D direction focused on real-time speech systems, language adaptation, training, evaluation and private deployment.
This is an expansion of Tender Touch's activity — not a replacement of its existing business.
Visit AUTO.MOTO.ptEuropean Portuguese is our first proving ground.
The project grew out of practical work on multilingual conversational voice systems. That work exposed a clear gap: many languages receive much less dedicated full-duplex model development than globally dominant languages, while sensitive organisations may also need alternatives to public-cloud-only voice platforms.
We chose European Portuguese as the first language programme to build and validate a repeatable process for language-specific training, evaluation and deployment.
From practical need to a working prototype.
Practical multilingual voice problem
Work on real voice-assistant systems revealed the need for better full-duplex support for less-served languages and private deployments.
Open-source technical baseline
A suitable full-duplex open-source architecture was selected as the starting point for experimentation.
European Portuguese data pipeline
Training experiments began using varied sources of authentic pt-PT speech, including podcasts, public audiovisual material and parliamentary speech.
Working pt-PT prototype
The project reached a functioning full-duplex European Portuguese prototype and moved beyond the initial feasibility question.
Continued training and evaluation
Training is ongoing. The principal constraint is now the speed of the R&D iteration cycle.
The bottleneck is now iteration speed.
The current training setup uses four GPUs. A substantial training cycle takes approximately one week on this configuration. That is enough to prove the pipeline works, but too slow for efficient research: each change to data, training strategy or evaluation can cost days of calendar time.
≈ one week for a substantial training cycle
Faster experimental iteration and comparison
Systematic experiments, evaluation and scaling
We do not claim linear scaling. Requested configurations are determined by model architecture, GPU type, memory, interconnect, dataset size and measured scaling behaviour.
More compute turns waiting time into research time.
More experiments
Test data and training changes without waiting weeks between decisions.
Better evaluation
Compare checkpoints and model variants systematically.
Better Portuguese
Spend more iterations on language-specific behaviour and quality.
Reproducibility
Repeat promising experiments and distinguish real gains from noise.
Private deployment
Study latency, model size and resource requirements for controlled environments.
Next language
Test whether the Portuguese adaptation methodology transfers to another European language.
Hear the current system.
Selected pt-PT demos will be published here with checkpoint dates and short technical notes.
Natural European Portuguese conversation
[audio/video placeholder]
Full-duplex / interruption example
[audio/video placeholder]
Longer dialogue + transcript
[audio placeholder]
Portuguese first. Then prove the method transfers.
A broader language R&D programme.
European Portuguese is the first language programme, not the end of the research direction. Once the training and evaluation methodology is validated, Tender Touch intends to assess additional European languages where high-quality real-time voice AI remains comparatively under-developed.
Future languages will be selected based on technical feasibility, data availability, access to native speakers and research partners, measurable need and deployment opportunities.
Replace placeholder email before publication.