An OpenAI TTS alternative for Hindi, priced in rupees

How do OpenAI TTS and Mira TTS differ?

OpenAI’s speech endpoint is one part of a much larger API. Its newest model, gpt-4o-mini-tts, takes a written instruction about tone and delivery, and the older tts-1 and tts-1-hd models are priced per character. The guide says the model follows Whisper’s language list, which includes Hindi, and also says the voices are optimised for English. If most of what you generate is English, or you need a dozen languages from one integration, that breadth is the reason to stay.

Mira TTS is the voice our own phone agents use, sold as a standalone API. We built it for Hindi and for the Hinglish that most Indian calls are in: a Hindi sentence that drops in “order confirm” or “refund process” without changing voice. It speaks the same request shape as OpenAI’s speech endpoint, so the decision is mostly about how your users sound and what you pay, and much less about rewriting code.

What does switching from OpenAI take?

Change two lines and add one field. Point the client at our base URL, set the model to mira-tts, pick one of our voices, and send response_format explicitly:

from openai import OpenAI

client = OpenAI(api_key="sk_live_...", base_url="https://sandbox.voice.miraiminds.co/v1")

with client.audio.speech.with_streaming_response.create(
    model="mira-tts",
    voice="ashu",
    response_format="pcm",
    input="चलिए numbers से start करते हैं।",
) as response:
    for chunk in response.iter_bytes():
        ...  # raw 16-bit mono PCM at 48 kHz, playable as it arrives

Our quickstart documents model, voice, input, response_format and sample_rate. Leave out OpenAI-only fields such as instructions rather than assume they do something. sample_rate is not an OpenAI parameter, so pass it through extra_body if you need 16 kHz or 8 kHz. Every response carries X-Chars-Billed and X-Cost-Paise headers, which makes it easy to check your bill against your own logs while you run both services side by side.

For a live conversation, one HTTP request per sentence pays for a new connection each time. Our WebSocket endpoint keeps one connection for the whole call and lets you cut a reply off when the caller interrupts. The OpenAI-compatible TTS guide walks through the first request in more detail.

What does it cost at volume?

Both publish a per-character price for at least some models. For 10 lakh characters (1,000,000), OpenAI’s tts-1 is $15.00 and tts-1-hd is $30.00. gpt-4o-mini-tts is billed on text input tokens and audio output tokens, so the cost follows the length of the audio you get back. Mira TTS is 1,000 × ₹1.80 = ₹1,800, before GST, and the output format or sample rate does not change that. A request that fails or stops part-way is not charged on ours.

If you are building a phone agent rather than buying a voice, the per-minute plans are the better comparison. Our agents are ₹1 a minute on T1 with a ₹1,00,000 monthly commitment, ₹3 a minute on T3 with ₹50,000 a month, or ₹5 a minute pay as you go, and that rate covers speech recognition, the model and the voice together.

How should you compare the two?

Write twenty or thirty lines your product will really say: customer names, rupee amounts, dates, an OTP prompt, and the English words people use inside Hindi. Generate each line on both services, in Devanagari where the sentence is Hindi, and play them to native speakers in random order without saying which is which. Then play the same lines over an actual phone call, because 8 kHz audio hides some problems and exposes others. The Hindi TTS comparison lists what the other Hindi voices on the market cost, with links to each price.