Real-time API · Early access
Real-time Romanian speech-to-text API for voice agents
Streaming transcription trained on how Romanians actually speak on the phone: compressed call audio, regional accents, and Romanian–English switching mid-sentence. Built by the team whose Romanian model ranked first on broadcast news and podcasts in an independent University of Bucharest benchmark.
In private development. Join the early access list.
- Caller
Bună ziua, aș vrea să verific statusul comenzii. Numărul e RO-48213.
- Agent
Sigur. Comanda RO-48213 e programată pentru livrare joi.
- Caller
Perfect. Și pot schimba adresa de delivery?
Trusted by
The problem
Generic speech models learn Romanian last
Phone-quality audio
Calls arrive compressed, at 8 kHz, with line noise and crosstalk. Models trained on clean podcasts lose words the moment audio hits a telephone line.
Romanian morphology
Romanian changes word endings for case, number and the definite article, so comandă, comanda and comenzii are different words. A model that gets the ending wrong gets the word wrong.
Romanian–English switching
Real callers say deadline, call, feedback and delivery mid-sentence. Single-language models stumble on the switch.
Names, numbers and IDs
Order numbers, addresses, product names and surnames are the words your agent needs most, and the words generic models miss most.
Where it fits
Built for the voice-agent loop
01
Stream the call
Send caller audio to the streaming endpoint as it arrives.
02
Read as they speak
Partial transcripts update word by word, and final transcripts lock in each segment.
03
Know when they're done
Turn detection signals the caller has finished, so your agent replies at the right moment.
In development04
Respond, then review
Your agent answers. After the call, the same API transcribes the full recording for QA and analytics.
Performance
Accuracy first. Measured on Romanian, then published.
We're training a new streaming architecture for Romanian, and the numbers improve every week. Before general availability we'll publish word error rate, end-to-end delay and head-to-head results against other real-time providers, all measured on the same Romanian audio, with the full methodology and its limitations. Until then, no headline figure goes on this page.
Word error rate (WER) is the share of words a system gets wrong; lower is better.
Capabilities
What works live, and what works on recordings
| Capability | Real-time | Recorded files |
|---|---|---|
| Romanian | In development | Available |
| English | In development | Available |
| Romanian–English code-switching | In development | Available |
| Partial and final transcripts | In development | Not applicable |
| Turn detection | In development | Not applicable |
| Speaker diarization | Planned | Available |
| Custom vocabulary | Planned | Available |
| Word timestamps and confidence scores | Planned | Available |
| Punctuation, capitalisation, numeral formatting | Planned | Available |
| PII redaction | Planned | Available |
| On-premise / private cloud | Planned | Available |
After the call
Same API, after the call
Production
Built for production traffic
Deployment
Cloud, private cloud or on-premise. Run the speech engine inside your own infrastructure when call audio can't leave it.
On-premise for real-time: plannedSecurity
GDPR compliant, ISO 27001 certified, SOC 2 Type II in progress. Built and operated in the EU.
Integrations
Plug it into your stack
- n8nComing soon
- MakeComing soon
Building on another voice or telephony stack? Tell us what you use; it shapes what we ship next.
Independent benchmark
Proven on Romanian, independently
Before real time, there was recorded audio. Researchers at the University of Bucharest benchmarked seven transcription systems on over 126 hours of Romanian speech. Vatis ranked first on broadcast news and on spontaneous podcast conversation.
Results on recorded audio, previous-generation model. Real-time results will be published separately.
| Model | In-domain | Out-of-distribution | ||||
|---|---|---|---|---|---|---|
| ProTV — Broadcast news | Antena 1 — Broadcast news | Audiobooks — Read literary prose | Films — Overlapping film dialogue | Stories — Children's stories | Podcasts — Spontaneous conversation | |
| Wav2Vec 2.0 | 27.7 | 25.6 | 40.8 | 75.4 | 54.1 | 36.3 |
| Whisper Small | 31.6 | 26.8 | 40.0 | 60.0 | 41.1 | 31.9 |
| Whisper Large | 12.3 | 5.9 | 14.8 | 27.3 — best in column | 10.9 — best in column | 11.7 |
| Whisper Small + Echo | 9.4 | 10.1 | 18.8 | 54.1 | 21.0 | 21.6 |
| Microsoft Transcribe | 2.9 — best in column | 4.8 | 10.6 — best in column | 31.1 | 17.6 | 11.5 |
| Google Chirp (USM) | 12.1 | 11.3 | 20.2 | 37.6 | 22.4 | 22.4 |
| Vatis | 5.2 | 4.4 — best in column | 13.0 | 31.2 | 16.0 | 10.2 — best in column |
Figures are Word Error Rate (%), lower is better. Outlined values are the best score in their column.
Early access
Get early access
Loading the form… If it does not appear, email contact@vatis.tech instead.
Frequently asked questions
Can't find the answer you're looking for? Reach out to our support team.
What is a real-time speech-to-text API?
A real-time (streaming) speech-to-text API converts speech into text while the person is still talking. Audio is streamed in small chunks and transcripts come back continuously, so applications like voice agents, live captions and call assistants can react during the conversation instead of after it.
Does Vatis support real-time transcription in Romanian?
Real-time Romanian is in private development, with early access opening to selected teams. Recorded-file transcription in Romanian and 50+ other languages is available today.
How accurate is real-time Romanian transcription?
We'll publish word error rate on Romanian audio, with full methodology, before general availability. Our current recorded-file model already ranked first on Romanian broadcast news and podcasts in an independent University of Bucharest benchmark.
What latency do voice agents need from speech-to-text?
Speech-to-text is one link in a chain: transcription, turn detection, the language model and speech synthesis all add delay. What callers feel is the total gap between finishing a sentence and hearing the reply. That's why transcription accuracy matters as much as speed: a fast transcript of the wrong words produces a fast wrong answer.
What is turn detection, and why does it matter for voice agents?
Turn detection decides when a caller has finished speaking. Too early, and the agent interrupts; too late, and the call feels slow. For Romanian it has to account for natural pauses, hesitations and trailing words. Turn detection for the Vatis real-time API is in development.
Can the same API handle live calls and recorded calls?
Yes. Use streaming during the call and recorded-file transcription after it for QA, compliance and analytics: one provider, one integration.
Does it handle Romanian and English in the same conversation?
Romanian–English code-switching works today on recorded files. For real time, Romanian and English are in training, with code-switching support rolling out as part of early access.
Can I run real-time transcription on-premise?
Vatis already deploys recorded-file transcription on-premise and in private cloud. On-premise deployment for real time is planned; tell us your requirements in the early-access form.
How do I get access?
Request early access with the form on this page, or book a call with the team.
Building a Romanian voice agent?
Tell us what you're building. We'll show you where we are and test on your audio first.