While much of today's artificial intelligence is text-focused, the market is moving quickly toward voice interfaces. Speech integration is emerging as the next big leap in customer experience, especially in complex flows like claims processing.
However, adopting this technology in Brazil brings a massive technical and cultural challenge. We need speech recognition systems (ASR) that understand our informality, spontaneous speech, and, above all, the country's vast diversity of accents.
Additionally, Brazilian Portuguese is full of words that sound the same but have different meanings — so-called homophones. This requires the AI to understand context so it doesn't make mistakes on critical data, like numbers and contract terms.
To find out which artificial intelligence actually delivers on this, our Research and Development (R&D) team conducted an in-depth study testing the best models on the market. The goal was to find out which one delivers the best accuracy, the highest speed, and the best cost-benefit ratio for our language, and we'll detail the main research findings throughout this article.
Choosing the ideal artificial intelligence model isn't just a matter of "which one understands better." The research concluded that the success of a transcription project depends on the balance between three factors.
The WER metric measures the system's word error rate. In practice, it answers: does the AI get regional accents right? Can it tell homophones apart, like "sessão" and "cessão" (session and grant)? The lower the WER, the more accurate and reliable the transcription is for your business.
RTF indicates whether the artificial intelligence can process audio faster than a person speaks. An RTF lower than 1 means the system is fast and ideal for real-time applications. If it's higher than 1, the tool is slower than human speech and will cause delays.
The decision here is between pay-as-you-go (cloud APIs) or having your own dedicated servers (self-hosting). The research proved that maintaining your own infrastructure only pays off financially at very high volumes, above 17,300 hours of audio per month. For the vast majority of companies, using APIs is considerably more efficient and economical.
The study carried out a systematic evaluation aimed at testing the stability and technical viability of these artificial intelligences in real-world scenarios, assessing everything from resistance to heavy noise to sensitivity to Brazilian accents.
The results brought surprising insights:
Gemini 2.0 Flash Lite is the clear champion in cost-effectiveness: Multimodal cloud APIs are financially unbeatable for the vast majority of operations. Google's model, for example, was up to 91 times cheaper than the best self-hosted solution.
Maintaining internal infrastructure (self-hosting) only pays off if your company transcribes very high volumes, above 17,300 hours per month.
A low overall error rate doesn't guarantee understanding of complex accents: Having good numbers on paper doesn't mean the AI understands the nuances of our language in practice. The accent from the state of Goiás was the most critical scenario in the test: almost all the artificial intelligences failed badly or "hallucinated" nonsensical text when trying to transcribe it.
The importance of cleaning audio with neural networks before transcription: The study proved that using AI to remove background noise before transcription (denoising) is vital for success. This smart technique cleans up the audio's noise while preserving voice characteristics, ensuring the model receives a clear sound and doesn't invent words where there was only noise.
The summary above covers only the main conclusions of the research. In the full paper, Anna Júlia de Souza Ferreira, our R&D team researcher, details:
At Tech for Humans (T4H), we design and implement fluid Digital Journeys and AI Agents.
As owners of our own technology, we don't rely on off-the-shelf solutions: we build custom projects to solve your business's specific challenges with the agility the market demands.
Major companies like Porto, Allianz, and MAPFRE have already anticipated this trend with us, replacing their old chatbots with true intelligent copilots capable of understanding, deciding, and executing complex tasks. The practical result is greater customer retention, higher operational efficiency, and a customer service experience elevated to a new level.
