Text to speech for voice agents, IVR and phone support

On a call, the gap between the caller finishing and the agent starting is the thing people notice. Chariot's streaming endpoint sends the first audio in about 75 ms (median, over WebSocket). The caller hears the reply begin before they have time to wonder if the line dropped.

The WebSocket endpoint speaks sentence by sentence. You can send text as your language model produces it, and the first sentence is already playing while the second is still being written. Each request takes up to 1,000 characters, which is more than one turn of a conversation needs.

The stream is raw 16-bit PCM at 44.1 kHz, mono, so there is nothing to decode before playback. If your telephony stack wants a different sample rate, resample on your side before it goes to the line. Speed is adjustable from 0.5x to 1.5x per request, which is useful when a caller asks the agent to slow down.

For fixed prompts such as greetings, menus and hold messages, POST https://api.chariot.in/v1/tts returns a WAV file you can store once and play on every call. There are Python examples to start from.

A support agent spends most of its time reading things that are not words: ₹850, order 2048, 06:45, 500 mg, a phone number, a URL. Chariot reads written text the way a person says it. An order number comes out digit by digit, a rupee amount comes out as an amount, and a time comes out as a time. You do not have to spell things out or pre-process the text before sending it.

In Chariot's own evaluation on English and Hindi, the model read hard written text (dates, currencies, URLs, units) correctly 94.0% of the time, first of the nine systems tested. The clip below is a Hindi payment confirmation: the amount and the word UPI are read as a person would say them.

This matters more on a phone than anywhere else, because the caller cannot see the screen and check. If the agent says the wrong amount, that is the amount the caller believes.

रिया को U P I से आठ सौ पचास रुपये भेज दिए गए हैं।

PlayPauseLoadingBufferingReplayRetry

0:04

Meera · Native Hindi

Most support conversations in India are not in one language. A caller starts in Hindi, asks "order confirm ho gaya?" and expects the reply in the same mix. Chariot reads Hinglish and other mixes of an Indian language with English as one sentence, without a break at the switch. It also reads Hindi written in Roman letters, which is how many scripts and chat transcripts arrive.

Chariot reads 15+ languages, including English, Hindi, Urdu, Bengali, Marathi, Telugu, Tamil, Gujarati, Kannada, Malayalam, Odia, Punjabi, German and Chinese. Hindi has 12 native voices and each of the other Indian languages has 2. Every voice can read every supported language, so one agent voice can serve a Tamil caller and a Hindi caller without changing character between them.

In the same evaluation, Chariot had the lowest Hindi character error rate (7.73%) of the seven systems tested in Hindi, and a naturalness score (NISQA) of 4.97 in Hindi, first among them.

नमस्ते! आपका order confirm हो गया है और कल शाम 5 बजे तक deliver हो जाएगा।

PlayPauseLoadingBufferingReplayRetry

0:06

Priya · Hindi voice

If your brand already has a voice, you can clone it. Record or upload 10 to 60 seconds of speech (about 30 seconds works best) and the voice is ready in about 10 to 20 seconds. A cloned voice speaks every language Chariot speaks, so a voice recorded in English can take Hindi and Tamil calls too. You must have the right to use any voice you clone, and consent is required.

Cloned voices come with paid plans: 3 on Starter, 10 on Creator, 50 on Startup and 500 on Scale.Cloned voices come with paid plans: 3 on Starter, 50 on Startup and 500 on Scale.

Concurrency is the other number that matters for a call floor. Starter allows 3 requests at the same time, Startup 10 and Scale 25. Pick the plan by how many calls you expect to be speaking at the same moment, not by total minutes.

Credits count generated audio: about 1,000 credits per minute of speech. A call where the agent speaks for one minute uses about 1,000 credits, whatever the call's total length.

Starter at ₹499 a month gives 100,000 credits, about 100 minutes, with 3 concurrent requests. Startup at ₹4,999 a month gives 1,500,000 credits, about 1,500 minutes, with 10 concurrent requests. Scale at ₹29,999 a month gives 10,000,000 credits, about 10,000 minutes, with 25 concurrent requests. Annual billing gives 2 months free.Starter at $5 a month gives 100,000 credits, about 100 minutes, with 3 concurrent requests. Startup at $49 a month gives 1,500,000 credits, about 1,500 minutes, with 10 concurrent requests. Scale at $299 a month gives 10,000,000 credits, about 10,000 minutes, with 25 concurrent requests. Annual billing gives 2 months free.

The Free plan gives 10,000 credits, about 10 minutes, with every voice and no card, which is enough to wire up the WebSocket and listen to your real prompts before you commit.

Text sent to Chariot and the audio it generates are processed in India. For a bank, an insurer or a telecom, that is usually the compliance team's first question.

Chariot is built by one of four teams selected under the IndiaAI Mission to build foundation models. The team behind Chariot has built voice work for HDFC Bank, Zomato and Doordarshan, among others, and more than 10,000 people use Chariot today.