Platform Release Notes

Voice Agents: Mid-call language switching — agents now detect and switch language within a single call without re-routing

STT: Bhojpuri and Haryanvi dialect models added — 31 languages now supported

Self-Hosted LLMs: Llama 3 Indic 70B INT4 model released — 40 tok/s on 2× A100 80GB

Hindi ASR WER improved from 3.8% to 3.2% on BFSI domain benchmark

Voice Agent E2E latency reduced from 750ms to sub-600ms

Fixed code-switching edge case where Hinglish utterances starting with English were incorrectly routed to English model

TTS: 15 new voice personas added — covering Bhojpuri, Maithili, and Rajasthani

API: Speaker diarization now available on batch and streaming endpoints (set diarize=true)

Platform: Voice Agent analytics dashboard v2 — real-time sentiment, CSAT, and resolution tracking

Tamil TTS naturalness score improved by 18% on MOS benchmark

Self-Hosted: Docker image size reduced by 40% for faster edge deployments

STT: Automatic PII redaction now supports UPI IDs, GSTIN, and vehicle registration numbers

Voice Agents: Freshdesk and Leadsquared CRM connectors added out-of-the-box

TTS: SSML <phoneme> tag support for custom brand name pronunciation

ONNX runtime upgraded to 1.17 — 12% throughput improvement on INT4 models

Resolved edge case where multi-digit OTP delivery in Telugu was misread at 0.5× speed

Self-Hosted LLMs: Gemma Indic 7B model released — optimised for edge deployment on RTX 4090

Platform: Breadcrumb and multi-step IVR flow builder in Voice Agent console

API: WebSocket ping/pong keep-alive for long-running streaming sessions

Kannada and Malayalam ASR accuracy improved by 15% with new training data

Voice Agents platform v2: LLM-powered conversation engine replaces rule-based intent routing

Self-Hosted LLMs: First release — Llama 3 Indic 8B and Mistral Indic 7B on Docker/K8s

STT: gRPC streaming endpoint launched — 15% lower overhead vs WebSocket for high-throughput pipelines

TTS: Voice cloning beta available under enterprise agreements

API: /v1/transcribe endpoint deprecated — migrate to /v2/stt before January 2026

Domain vocabulary adaptation: Healthcare (ICD-10, SNOMED) and Legal domain models released

TTS: Emotion and tone control via API parameters (professional, empathetic, urgent)

STT streaming latency reduced from 120ms to sub-90ms first-token delivery