Skip to content

Streaming Hinglish text-to-speech

One API, two production voices, first audio in tens of milliseconds, at 20 paise per audio-minute. Hindi, English and the code-mixed Hinglish that Indian customers actually speak, with a normaliser that reads amounts, dates, OTPs and phone numbers the way a trained agent does.

41 msend-to-end time to first audio, warm, single stream, Mumbai (c3-standard-4)
0.14–0.19real-time factor per stream: a second of speech in under 200 ms of compute
12.4×audio-seconds per wall-second from one 4-vCPU node at four concurrent streams
₹0.20per audio-minute, metered per second, same price for REST and WebSocket

QuickstartCreate a key, make the first call with curl, Python, Node or a browser. Streaming guideChunked REST vs WebSocket, the event timeline, buffering and playback. API referenceEvery field, status code and event, straight from the server source. Text normalisationWhat the Hinglish normaliser does to numbers, acronyms, dates and OTPs, with the 66 canonical cases. VoicesAnjura and Raj, with samples; how to limit a key to a voice. PricingOne number. How metering works and what a cancelled stream costs.

The sixty-second version

export VARTALAAP_API_KEY="vt_live_…"   # from https://tts.vartalaap.si/keys
curl -sN -X POST https://api.vartalaap.si/v1/tts \
  -H "x-api-key: $VARTALAAP_API_KEY" -H "content-type: application/json" \
  -d '{"text":"आपकी EMI ₹12,450 है, जो 5 नवंबर को due है.","voice":"anjura","format":"wav"}' -o emi.wav

The response body is a WAV stream: the first bytes arrive after the first 80 ms frame is decoded, so a player can start before synthesis finishes. For conversations, open a WebSocket once and send sentences as your LLM produces them; the server tells you the exact moment of first audio (first_audio_ms) and the real-time factor for every utterance.

How the engine behaves

  • Normalises first. ₹12,450 → बारह हज़ार चार सौ पचास रुपये, EMI → E M I, 5 नवंबर stays a date, OTP 482913 is read digit by digit. Preview it without synthesising.
  • Chunks on sentences, at most 50 tokens each, and streams chunk after chunk on one connection.
  • Caches text prefills per (voice, chunk): a repeated IVR line starts in ~18 ms of engine time.
  • Says no cleanly. When every worker is busy you get 503 with Retry-After (or an error event with retry_after), nothing is billed, and you retry. Limits & errors.

Where things live
API https://api.vartalaap.si (REST + WebSocket), Mumbai
Developer portal tts.vartalaap.si: keys, playground with live TTFA, usage, billing, voice requests
Status tts.vartalaap.si/status, GET /readyz
Changelog changelog
Company vartalaap.si