Speech-to-Text and TTS for Indian Languages: API Guide

Why general-purpose speech models fall short for Indian languages

Whisper transcribes English (and other major world languages) well, but accuracy drops noticeably on Indian-accented English and falls further on native Indic languages. If your product needs Hindi, Tamil, Telugu, or Bengali voice input — or natural-sounding Indic-language speech output — a model built for those languages performs meaningfully better than a general-purpose one stretched to cover them.

Sarvam AI is an India-first voice provider built specifically for Indian languages. Through AICredits, you call it with the same OpenAI-compatible audio endpoints you'd use for Whisper or OpenAI's TTS — same API shape, same INR wallet, no separate account.

Text-to-speech in Hindi

curl https://api.aicredits.in/v1/audio/speech \
  -H "Authorization: Bearer sk-your-key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "sarvam/bulbul-v2",
    "input": "नमस्ते, आप कैसे हैं?",
    "voice": "anushka"
  }' \
  --output speech.wav

Sarvam's TTS voices are called speakers and use Indian names — anushka, rahul, priya, karun, and more (44 speakers across 12 languages in total). If you already have code using OpenAI's voice field (alloy, nova, shimmer), AICredits maps each one to a Sarvam speaker automatically, so a codebase built for OpenAI TTS can point at sarvam/bulbul-v2 with no other changes.

Speech-to-text for Indian languages

import base64
from openai import OpenAI
 
client = OpenAI(base_url="https://api.aicredits.in/v1", api_key="sk-your-key")
 
with open("customer_call.wav", "rb") as f:
    audio_b64 = base64.b64encode(f.read()).decode()
 
response = client.audio.transcriptions.create(
    model="sarvam/saarika-v2",
    file=("customer_call.wav", open("customer_call.wav", "rb"), "audio/wav"),
    language="hi-IN",  # or ta-IN, te-IN, bn-IN, mr-IN, gu-IN, kn-IN, ml-IN, pa-IN
)
print(response.text)

sarvam/saarika-v2 is tuned for Indian accents and code-mixed speech (Hindi-English switching mid-sentence, common in real customer calls) — a pattern that trips up general-purpose transcription models trained mostly on monolingual data.

Supported languages

| Code | Language | Code | Language | |------|---------|------|---------| | hi-IN | Hindi | ta-IN | Tamil | | en-IN | English (India) | te-IN | Telugu | | bn-IN | Bengali | mr-IN | Marathi | | gu-IN | Gujarati | pa-IN | Punjabi | | kn-IN | Kannada | ml-IN | Malayalam |

When to use Sarvam vs. Whisper/OpenAI TTS

| Use case | Better choice | |----------|----------------| | English-only voice UI | openai/whisper-1, openai/tts-1 | | Hindi, Tamil, Bengali, or other Indic-language voice UI | sarvam/saarika-v2, sarvam/bulbul-v2 | | Code-mixed customer calls (Hindi-English) | sarvam/saarika-v2 | | Multilingual product with English as primary | Either — test both against your actual audio samples |

Both routes run through the same AICredits key and INR wallet, so switching between them for A/B testing is a one-line model change, not a new integration.

Frequently Asked Questions

Do I need a separate Sarvam account?

No. AICredits routes sarvam/ model IDs directly to Sarvam's API — you only need your AICredits key.

Can I use Sarvam models for a WhatsApp voice-note bot?

Yes. Download the voice note via Meta's media API, then transcribe it with sarvam/saarika-v2 for Indian-language accuracy before passing the text to an LLM for a reply. See Build a WhatsApp AI Bot for Your Business for the full flow.

What does Sarvam cost compared to Whisper?

Sarvam pricing is INR-native and billed per request or per second of audio depending on the model; Whisper is billed per minute in USD-converted rupees (~₹0.55/minute). Check aicredits.in/models for the current per-model rate, since both are usage-based and can vary by request size.

Can I mix Sarvam TTS with an OpenAI or Claude model for the text generation step?

Yes — generate the reply text with any chat model, then pass that text to sarvam/bulbul-v2 for speech output. All three calls (LLM, TTS, STT) draw from the same wallet.

Get started

• Read the audio API docs for the full parameter reference and voice mapping table.
• Create a free account and test Sarvam's Hindi TTS with your own text.
• Related: Build a WhatsApp AI Bot for Your Business.