Koyal
Intro
Koyal is Plivo’s in-house text-to-speech model, launched on 18 November 2026 for voice AI agents that need low latency. The official post Introducing Koyal describes it as Plivo’s first owned TTS model, built so quality, speed, and price do not have to be traded off one against another.
The post lists 14 languages: Assamese, Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Nepali, Odia, Punjabi, Spanish, Tamil, and Telugu. Plivo’s published time to first audio is under 200 ms. It says Koyal is hosted in India and the United States, and that audio is processed in the region you choose, to keep latency down and help with data-residency rules. The product page is Koyal text to speech API.
This entry records the new model name Koyal. It should not be merged with Plivo’s existing communications APIs, or with other TTS products such as ElevenLabs, Cartesia, or Fish Audio.
Advertisement
Explain
Plivo’s stated context is scale: it carries more than 5 billion voice calls a year, and more of those calls are now answered by voice agents. Teams kept running into the same tradeoffs: TTS that sounded good was often slow, expensive, or shaky when it had to switch languages. Koyal is positioned as a first-party model for live calls, with fast first audio, local reading of amounts and codes, and a lower list price billed by the minute.
On price, the blog lists $0.0065 per minute in the US and ₹0.65 per minute in India. The comparison table treats 1,000 characters as one minute of speech and, using pay-as-you-go list prices from 18 November 2026, says Koyal is more than 65% lower than several widely used models in the table. Latency figures come from Coval’s 30-day average or from providers’ own pages; Plivo notes they are not all measured the same way, and Koyal’s 200 ms is Plivo’s own figure.
Plivo also highlights how Koyal reads speech in context: for example, ₹1,25,000 as “one lakh twenty-five thousand rupees,” and card numbers and dates spoken clearly rather than digit by digit. The docs in Text-to-Speech Overview describe batch mode (POST text, get a whole audio file) and realtime mode (stream frames over a WebSocket, with up to eight replies on one socket), plus telephony formats such as 8 kHz μ-law and A-law. The language table in the docs may not match the 14-language list in the launch post exactly; use the live docs when you integrate.
Usage
Koyal is available to every Plivo account. Start with Introducing Koyal and the product page Koyal text to speech, then follow Text-to-Speech Overview to choose batch or realtime. If you already run voice agents on LiveKit, Pipecat, or your own stack, Plivo says you can point the agent at Koyal without changing the rest of the orchestration.
When you integrate, pick the US or India region, a voice that matches the language you need, and 8 kHz μ-law or A-law if the audio must go onto a phone call. For stable readings of names and codes, the docs support custom pronunciations and lexicons (IPA, ARPAbet, and respelling). Realtime streaming fits an agent that talks as the LLM generates text; batch is a better fit for pre-recorded prompts and notifications.
When you evaluate latency and price, measure time-to-first-audio and the bill on your own text, language, and region, rather than relying only on the blog’s comparison table. Koyal is a TTS product, not a new general-purpose foundation model.