Change Log

Call recordings are now captured in stereo, with the user on the left channel and the agent on the right. Mono recordings mixed both speakers onto a single channel, so overlapping speech was summed and could clip; in stereo each speaker stays isolated, which makes recordings easier to listen to and better suited for downstream processing (transcription, analytics). This is now enabled for all calls. See Call recordings.

Bulbul V3 reaches end of life on Voice Agents on November 10, 2026. Switch every agent that speaks with Bulbul V3 to Bulbul V4 before that date. In the agent’s Settings, set the model under Speaking to Bulbul V4 and pick a voice. To keep a voice you use today, find its Bulbul V4 name in Voice name changes from V3 to V4. This change applies to Voice Agents only. Bulbul V3 stays available on the Text-to-Speech API.

You can now control how loud your voice agent sounds. A new Volume slider in the agent’s voice settings (Sarvam / Bulbul voices) sets the loudness of the agent’s speech, from 0.7 (quieter) to 2.0 (louder), with 1.0 as the natural level. The volume is stored with the agent’s voice configuration and applied on every call, so it can be set once and manipulated later. See Speakers & voice.

The sarvamai Python and TypeScript SDKs now expose Instant Voice Cloning (beta) with the same call shape as Bulbul Text-to-Speech: voice_cloning.convert / voiceCloning.convert for one-shot synthesis, convert_stream / convertStream for HTTP streaming, and voice_cloning_streaming / voiceCloningStreaming for WebSocket sessions (configure → convert → flush). Voice lifecycle helpers (create_voice, list_voices, get_voice, delete_voice) live on the same voice_cloning client. Paths stay under /voices/*. Requires the text_to_speech_voice_cloning capability. See the Instant Voice Cloning overview and install sarvamai ≥ 0.1.37 (Python) or sarvamai ≥ 1.1.12 (TypeScript). Older alphas such as 0.1.31a7 do not include the full streaming APIs.

Language model prices have changed. sarvam-105b and sarvam-105b-conversations now cost ₹15 input, ₹5 cached input, and ₹60 output per 1M tokens. gemma4 is now ₹35 / ₹10 / ₹90. glm5.3 cached input and output drop to ₹20 and ₹390, and input stays at ₹126. deepseekv4.1-flash is now ₹25 / ₹0.5 / ₹105. See Pricing for the full table.

POST /v2/chat/completions now accepts an optional x-session-affinity header on the gemma4 model. Send the same session ID on every turn of a conversation, and later turns can reuse cached prompt context for lower latency. Session IDs are scoped to your API subscription, and requests without the header work as before. See Session affinity.

Bulbul V4 is now available for Voice Agents, alongside Bulbul V3. The voice picker is now a filterable catalog: browse and preview voices, filter the list to find the right fit, and choose between Bulbul V3 and Bulbul V4 per agent. See Speakers & voice.

Voices in Bulbul V4 carry new, more descriptive display names: Shubh is now Shubh - Hinglish Ad Voice and Ritu is now Ritu - Hindi Support Agent. The full V3-to-V4 name mapping is in Speakers & voice.

Extract is now more accurate across Indic, English, and mixed-language documents, with lower latency and support for longer documents. Every extracted value now includes a citation to its source, viewable in the UI, so each field can be verified at a glance. Extract rate limits have also been increased. Get started with the Document AI overview.

Instant Voice Cloning (beta) now supports streaming audio. Use POST /voices/clone/stream for a single binary HTTP stream, or GET /voices/clone/ws for a persistent WebSocket that accepts incremental text. See the HTTP streaming and WebSocket guides, or the API Reference.

The Dubbing API is now available for programmatically localizing audio and video while preserving each speaker’s voice. Learn how to create and manage dubbing jobs in the Dubbing API documentation, or explore the endpoints in the API Reference.

The Voice Cloning API is now available for generating speech in a cloned voice from a short reference recording. Get started with the Voice Cloning API documentation, or explore the endpoints in the API Reference.

Workspace owners can now invite people directly from the workspace members page, with the workspace pre-selected, instead of going through organisation-level settings.

Connecting an agent to Slack no longer means creating a Slack app by hand and pasting secrets into a form. Copy one command into your Slack workspace, approve the permissions, and you are done - no Slack dashboard required.

Added live voice input: words appear on screen while you are still speaking, instead of recording a clip and waiting for it to be transcribed. The 30-second limit is gone; you can speak for up to five minutes in one go.

You can now share any file your agent creates with a link - from the chat card, the Files panel, or the preview. Choose who can open it (your workspace, your organisation, or anyone with the link) and change the audience later without breaking the link. Recipients see a clean page with just the file, and public links need no account.

Agents now finish multi-step jobs faster by batching their work into a single pass instead of one step at a time. Long results from sub-tasks, such as a full report or a long rewrite, now arrive complete instead of being cut off and redone from scratch.

keyterms is now supported on the streaming WebSocket endpoints — /speech-to-text/ws (legacy) and /speech-to-text-realtime/ws (realtime) — with model="saaras:v4", matching the existing REST and Batch behaviour. Pass a JSON-encoded list of up to 50 terms (64 characters each) as the keyterms query parameter when opening the connection. See Keyterm Prompting.

Fixed an issue where agents on a schedule, such as a daily report, posted their text but dropped the files they created. Files now arrive with the message, every time. See Scheduled Tasks.

You can now change a specific part of an artifact your agent built by commenting on it directly. Turn on Comment in the preview, click the block or button you want changed, and type your feedback. Each comment attaches to that element and reaches the agent with your next message, so you no longer have to describe where the problem is.