Kognivya cloud
Fastest start. Hosted in India-region data centres, with usage-based billing.
Kognivya Technologies
How streaming text-to-speech works
A request travels through five stages. Every stage is built to hand work to the next as early as possible.
The pipeline
Text arrives in pieces
Your app or language model sends text as it is produced, over a WebSocket or HTTP stream.
Normalise
Numbers, dates, currency and acronyms are expanded the way an Indian English speaker would say them: "₹4,50,000" becomes "four lakh fifty thousand rupees".
Generate speech in chunks
The acoustic model and vocoder work on short phrases, so audio for the first phrase is ready before the sentence ends.
Stream audio and timestamps
Audio chunks and word timings flow back immediately. Your player starts, and captions stay in sync.
Handle interruptions
If the listener speaks, your app sends a cancel message and playback stops within one chunk.
Latency budget
Design targets for a short sentence on a nearby region. We publish measured numbers when the beta opens.
| Stage | Target |
|---|---|
| Network to nearest region | 30 to 60 ms |
| Normalisation | under 10 ms |
| First audio chunk generated | 150 to 200 ms |
| Network back to client | 30 to 60 ms |
| Time to first audio | about 300 ms |
Deployment options
Your cloud account
We deploy into your Azure, AWS or Google Cloud tenancy so data never leaves it.
On-premises (planned)
For intranet-only environments, with all models running on your own servers.