Ragini Lite: 100× Real-Time Code-Switched TTS for Scale & Edge

Text-to-speech has to do more than sound good in a quiet demo. In production, it must pronounce names, dates, amounts, account numbers, and mixed Hindi–English sentences correctly. It must also respond quickly when thousands of conversations are competing for compute.

Ragini Lite is our answer to that operational challenge: a compact Hindi–English text-to-speech model packaged as a production-ready NVIDIA Triton service. It is designed for the short, high-frequency utterances used by customer-support agents, IVRs, payment reminders, delivery updates, collections workflows, and other enterprise voice applications.

In our deployment benchmark, Ragini Lite generated audio at roughly 118× real time on GPU — on the order of 100 seconds of speech for every second of inference. On CPU it ran at nearly 16× real time. That efficiency makes high-volume speech generation possible without the infrastructure footprint normally associated with larger generative speech models.

From Ragini to Ragini Lite: Efficiency by Design

At ConvoZen Research, our voice synthesis work began with Ragini — our high-fidelity engine for complex Indian-language code-switched environments like Hindi–English and Tamil–English. Ragini set the quality bar for expressive, natural speech across scripts.

But the best model for a voice agent is not necessarily the largest one. A great many enterprise utterances are short and functional; they need to be clear, consistent, and available immediately, at a cost that scales to millions of calls. Ragini Lite was built directly for that workload: the same commitment to natural Hindi–English speech, re-engineered around latency and economics.

The Cost Problem in Production TTS

A production system has to balance several constraints at once:

Large speech models can produce expressive audio, but serving them for every confirmation, reminder, OTP, balance statement, or IVR prompt can be unnecessarily expensive. Many enterprise utterances are short and functional. They need to be clear, consistent, and available immediately.

Ragini Lite is optimized for exactly this part of the workload. It synthesizes a large amount of audio from a small amount of compute, running faster than real time even on CPU, while GPU deployment provides additional headroom for latency-sensitive or high-throughput applications.

Built for Hindi, English, and the Space Between Them

Real Indian customer conversations rarely stay inside one language. A sentence may begin in Hindi, switch to an English product term, contain an amount or date, and end with a Hindi instruction:

Naive text-to-speech pipelines often treat such text as an edge case. Ragini Lite treats it as a normal input pattern — Hindi and English in the same sentence, along with the numbers, dates, and amounts that fill real conversations, handled as ordinary input. The model produces a 22.05 kHz waveform in a single inference pass.

A Compact Model in a Production Serving Stack

Ragini Lite is served as a portable inference graph through an NVIDIA Triton service:

Input text
    │
    ▼
Neural acoustic inference — ONNX Runtime (CPU or GPU)
    │
    ▼
22.05 kHz audio waveform

The service exposes the pipeline as a single endpoint, which keeps behavior consistent across clients and makes the deployment easier to operate, scale, and monitor. ONNX Runtime provides a portable, hardware-agnostic inference graph, while Triton supplies production serving primitives such as HTTP/gRPC endpoints, model instances, metrics, and pipeline orchestration. The GPU configuration runs two model instances for additional throughput headroom.

Voice behavior can be adjusted at runtime through three familiar controls:

These settings favor stable timing and clean delivery — useful properties for transactional and conversational voice applications.

Performance: Much Faster Than Real Time

We benchmarked the end-to-end Triton pipeline across 50 varied requests spanning short, medium, and long Hindi, English, and code-mixed utterances from domains such as banking, travel, healthcare, e-commerce, and IVR.

Because synthesis runs far faster than playback, the system generates roughly 100 seconds of speech for every second spent in inference. Every tested configuration remained faster than real time. GPU offers the best interactive latency and throughput headroom, while CPU remains a viable option when infrastructure cost is the dominant constraint.

The GPU path is roughly 7× faster than CPU, generating well over 100 seconds of audio per second of synthesis time. An alternative accelerated-runtime experiment produced similar latency but required substantially more GPU memory, so ONNX Runtime remains the more efficient serving choice for this model.

Why This Efficiency Changes the Economics

Synthesis speed is more than a convenience. It is a direct signal of how much compute is required to produce a given duration of speech.

On GPU, one second of inference corresponds to roughly 118 seconds of generated audio; on CPU, to nearly 16 seconds. Actual production throughput and cost will also depend on request distribution, hardware, utilization, networking, and orchestration overhead, but the benchmark shows why Ragini Lite is inexpensive to operate: it spends far less compute time synthesizing audio than users spend listening to it.

This makes Ragini Lite especially useful for:

Small by Design, Production-Ready by Construction

Ragini Lite is not intended to replace every expressive or creative speech model. It is built for the much larger set of moments where TTS needs dependable pronunciation, low latency, and excellent unit economics.

By combining a compact Hindi–English voice with portable ONNX inference and a Triton serving pipeline, Ragini Lite delivers a pragmatic foundation for enterprise speech: fast enough for real-time applications, efficient enough for high-volume workloads, and simple enough to deploy within an organization's own infrastructure.

Ragini Lite makes natural, code-mixed voice generation affordable at the scale where TTS becomes useful.

Bringing Intelligence to the Edge

Our vision with Ragini Lite has been to bring intelligence where the customer demands it — where hardware is not the limitation. We are working closely with partners who want to build global-standard TTS provider capabilities for India use cases on their own edge infrastructure.

At the moment, we are not releasing Ragini Lite on API access. However, on our developer portal you will be able to use it and try it via our playground.