Text-to-Speech

Usage Instructions

Generate natural-sounding speech from text.

Sanjaya is Ayantram's own hosted speech service and the default provider. It needs no API key, speaks 600+ languages including every major Indian language, supports emotion presets, and can use a voice cloned for your organization. Usage is billed to your workspace at ₹15 per 10,000 characters (about $0.158), charged on the characters you submit.

The other providers — OpenAI, Deepgram, ElevenLabs, Cartesia, Google Cloud, Azure and PlayHT — are billed by those vendors to your own API key.

Language is the script you write

There is no language setting to get right. Write Devanagari and the voice speaks Hindi; write Latin and it speaks English; write Bengali and it speaks Bengali — all in the same voice, and all in one line if you mix them:

Hey, just do me a favor, मेरी बात सुनो, বাকি সব ঠিক হয়ে যাবে।

That renders as one continuous voice with no seams.

Numbers are spoken correctly by default

"Speak digits as words" is on by default and should stay on. The engine mis-reads bare digit runs in Indic scripts — डोलो 650 मिलीग्राम has been rendered as "105", "605" and "750" — so a dosage, price or OTP read from digits is a real-world error, not a cosmetic one. With the setting on, the number is spelled out before synthesis and read correctly.

The service also tells quantities from identifiers using context: कुल ₹320 is read as an amount, while आपका ओटीपी 4400 है is read digit by digit ("डबल चार डबल शून्य"), because a customer cannot type back "four thousand four hundred".

Tools

tts_sanjaya

Convert text to speech using Sanjaya, Ayantram's hosted multilingual service. No API key required.

Input

Output

tts_openai

Convert text to speech using OpenAI TTS models

Input

Output

tts_deepgram

Convert text to speech using Deepgram Aura

Input

Output

tts_elevenlabs

Convert text to speech using ElevenLabs voices

Input

Output

tts_cartesia

Convert text to speech using Cartesia Sonic (ultra-low latency)

Input

Output

tts_google

Convert text to speech using Google Cloud Text-to-Speech

Input

Output

tts_azure

Convert text to speech using Azure Cognitive Services

Input

Output

tts_playht

Convert text to speech using PlayHT (voice cloning)

Input

Output

Streaming speech

The block returns a finished audio file. For a live conversation — a voice-mode chat, or an agent whose reply should start being spoken before it is finished — use the streaming endpoint instead, which returns first audio in roughly half a second regardless of how long the rest of the sentence is.

POST /api/proxy/tts/sanjaya/stream accepts two request shapes:

One line, spoken and closed (Content-Type: application/json):

{ "text": "आपका ऑर्डर कन्फ़र्म हो गया है।", "voiceId": "sanjaya:aarohi" }

Text streamed in as it is produced (Content-Type: application/x-ndjson) — the shape to use when piping a model's tokens straight into speech. The first line configures the session; every line after it is either text or a control action:

{"voiceId": "sanjaya:aarohi"}
{"text": "Hello there, "}
{"text": "your order is confirmed."}
{"action": "end"}

The socket to the speech service is opened once for the whole reply rather than once per sentence, and each completed sentence is released for synthesis as soon as it lands — so audio keeps flowing while the model is still writing.

The response is raw 16-bit PCM, not a WAV file: a container needs a length in its header, which would mean buffering the whole clip and losing the latency the endpoint exists for. Read the format from the response headers rather than assuming it, because a telephony caller gets 8 kHz G.711 from the same endpoint:

Streamed speech is billed the same way, on the characters sent.

Choosing a voice

Every voice available to your workspace lives under Settings → Tools → Voices, on two shelves: Preset voices, the personas we ship and maintain, and Your voices, the ones your organization has cloned.

Each row has a play button, so you can hear a voice before you put it in front of customers. By default it speaks a standard sample line — but the box above the list takes your own line, and every voice then reads that instead. This is the useful version of an audition: whether a voice sounds pleasant reading a greeting we wrote tells you much less than whether it can say your product names, your customers' names, and your actual script without sounding wrong.

Cloned voices

A cloned voice is one your organization registers from a short recording of a real person. It is visible and usable only inside your own tenant, it appears under Your voices, and it can be removed at any time — which is also how consent is withdrawn.

Cloning a voice

An organization admin can add a voice in about a minute. 10-20 seconds of one person speaking is all it takes, and the cloned voice then speaks any language exactly the way a preset does.

The new voice appears in the same list, tagged Organization, and is immediately selectable in the block's Voice field and by the tts_sanjaya agent tool.

Recording in the browser

The recorder is tuned for cloning rather than for a phone call, which is where most of the quality difference comes from:

• Every voice-processing feature is switched off. Browsers enable echo cancellation, noise suppression and automatic gain control by default. All three are actively harmful here — gain control flattens the dynamics that make a voice recognisable, noise suppression carves holes straight through it, and echo cancellation can duck the speaker mid-sentence. The engine clones what it is given, so it would learn all three artifacts.
• The audio is never compressed. Samples are captured straight off the audio graph and written as a 16-bit WAV, rather than the Opus the browser's own recorder produces. Opus thins exactly the top octave the cloning engine is most sensitive to.
• A passage is provided to read, in English or Hindi. The reference sets the clone's accent and pace, not just its timbre, so reading a natural customer-facing passage produces a better voice than talking at random.

Before anything is uploaded, the clip is checked in the browser and you are told what is wrong with it:

A cloned voice is visible and usable only inside the organization that uploaded it. The built-in preset personas are shared by everyone.

Adding a voice takes a few seconds because the new voice has to prove it can speak its own reference text before it is accepted — an unusable recording is rejected at upload rather than failing later in a live workflow.

Notes

• Category: tools
• Type: tts