Live Portuguese transcription across channels
Portuguese speech to text for voice agents works for someone tapping a mic button in a banking app, talking to a car or asking a website for help. Hear accepts a live PCM16 stream over a WebSocket and hands back the first partial in about 200 ms. PyAI Omni, the speech to speech agent, uses Hear underneath, so you can run Hear with your own logic and voice or let Omni handle listening, reasoning and speaking in one session. Live Portuguese transcription from that stream can also drive captions. Set the language to Portuguese on the request, add up to 5 custom vocabulary terms for names such as Pix, Nubank or a street in São Paulo, and the same stream serves a support line, a kiosk or an in-car assistant without a second integration. The same engine also works as a Portuguese speech to text generator for the transcripts your team keeps after each session.
Key takeaways
- First partial: Portuguese text back in about 200 ms.
- Accuracy, English measured: 1.59% WER clean and 4.76% on 8 kHz phone, accented or noisy audio, PyAI benchmarks at industry parity, with no Portuguese score published.
- Channels: apps, browsers, cars, devices and lines share the same WebSocket stream.
- Price: $0.001/min by the whole minute, no platform fee.
- Inside Omni: turn end about 196 ms, voice to voice near 390 ms median.
- Compatibility: whisper-1 compatible through a base URL change.
From spoken Portuguese to a finished task
Banking app, car or browser, the route from voice to words is identical, so one integration serves every surface.
A person in São Paulo or Lisbon talks into an app, a browser, a device or a line.
The client sends PCM16 frames of about 20 ms over a WebSocket while the speech continues.
Hear sends a first partial about 200 ms in, such as fazer um Pix, and refines it word by word.
Your agent, or Omni, reads the final text and sends the transfer, shows the balance or answers.
Portuguese speech realities Hear works with
Two main varieties, nasal sounds, spoken IDs and borrowed words all show up in Portuguese voice input.
Brazil and Portugal
Recife, São Paulo, Porto and Lisbon speakers sound quite different. Test on recordings from your own users, because PyAI publishes no Portuguese score.
Nasal vowels and reduced syllables
Sounds like ão and the clipped syllables of European Portuguese are worth checking on fast speech samples before you launch.
CPF, CEP and amounts
With numerals handling, CPF numbers, CEP postcodes and money amounts return as digits ready for a form or a transfer.
App names and English words
Portuguese users mention English app and brand names. Up to 5 custom vocabulary terms make sure those land spelled correctly.
Street and traffic noise
Buses, motorbikes and busy shops add background sound. Hear measures 4.76% WER on noisy and accented audio, measured on English.
Early partials, clean logs
Partials begin while the user talks, and verbose JSON returns speaker turns and timestamps for audit trails.
Accuracy and speed
Published accuracy is PyAI's own English measurement at industry parity. Take it as a guide for Portuguese and review your own recordings. Omni timing figures come from full Omni sessions.
Portuguese jobs for a listening assistant
Portuguese voice assistant speech to text fits any product where Brazilian or European users prefer speaking. Six jobs, equal in weight:
Banking and payment apps
Let users ask for a balance or start a transfer by voice, then confirm the amount on screen.
Delivery and ride apps
Turn a spoken address or order note into text so a courier or driver gets it right.
In-car assistants
Handle navigation and messages for drivers in Brazilian traffic without them touching a screen.
Field service logging
Let technicians speak job notes and meter readings on site, saved as text and digits.
Voice tutor apps
Check a learner's spoken Portuguese against the target phrase and point out what was missed.
Customer support phone line
Answer a Portuguese help number, understand the caller's issue and send it to the right team.
Formats, limits and output
| Input | Portuguese PCM16 audio over a live WebSocket from an app, browser, device or line, or one file up to 25 MB. whisper-1 clients connect unchanged. |
|---|---|
| Output | Partial text during speech and a final transcript after, delivered as text, JSON or verbose JSON with timestamps and speaker turns. |
| Modes | Sync file call for short clips, live streaming for apps, devices and voice sessions, and background jobs for long recordings that run 92 to 247 times real time. |
| Tuning | Set Portuguese as the language, turn on numerals handling for CPF numbers and amounts, and add up to 5 custom vocabulary terms such as app or bank names. |
| Billing | $0.001/min of audio, rounded to the whole minute. An empty transcript costs nothing, and a busy engine returns a retry, not a charge. |
Wire Hear into a Portuguese assistant
Hear follows the OpenAI transcription format, so a Portuguese assistant already on whisper-1 moves over with a new base URL. Docs: PyAI SDKs.
SDK
In your app backend, the SDK opens a stream for each Portuguese session and delivers every partial and final transcript to your logic as Hear produces it.
MCP
Through the MCP server, an agent can call Hear as a tool to transcribe a Portuguese audio message from a user before replying.
CLI
Use the CLI on a Portuguese recording to see partials print live, a fast way to check bank or product names before going live.
Where Hear fits
PyAI is voice infrastructure. Hear, Speak and Omni run in one region on one stack, so a Portuguese session moves from speech to reply without vendor hops. Hear is not an assistant designer and does not write your dialogue. Numbers on PyAI benchmarks were measured on English, so pilot with real Portuguese speakers first. Run that pilot on sandbox keys, which never bill, and compare the transcripts against your own call recordings before switching live traffic over.
Portuguese voice agent questions
How quickly does Hear return Portuguese text?
Hear delivers the first Portuguese partial about 200 ms after the user starts speaking and keeps correcting it as more audio arrives. Running under Omni, the end of a turn registers roughly 196 ms after the person goes quiet, and the spoken answer follows at around 390 ms median voice to voice.
Has PyAI measured Portuguese accuracy?
No. PyAI publishes Hear results of 1.59% on clean audio and 4.76% on 8 kHz phone, accented or noisy audio, all measured on English at industry parity. A team building for Brazil or Portugal should transcribe real sessions from its own users and compare them with what was actually said.
Can I add Hear to a Portuguese assistant I run today?
Yes. Open a WebSocket per session, stream PCM16 audio from your app, browser or device, and pass Hear's partial and final text into your own logic and voice. Assistants using the whisper-1 compatibility alias usually need only a new base URL, and output comes as text, JSON or verbose JSON.
How is Portuguese transcription priced?
Hear charges $0.001 per minute of audio, billed by the whole minute, with no platform fee, and empty transcripts are free. A new account includes $50 in credit with no credit card required, and sandbox keys never bill, so a banking or delivery flow can be tested end to end first.
Should a Portuguese product use Hear or Omni?
Use Hear when your team already owns the assistant logic and voice and needs quick Portuguese transcripts. Use Omni when PyAI should listen, spot the end of each turn, reason and speak in one session, with a turn detector that waits through half finished thoughts and spoken numbers.
Prices checked November 2026. Hear figures from pyai.com/benchmarks.