When a service has two WebSocket speech interfaces, a naming error can break live subtitles. Saaras v3 Realtime was released for a different interaction pattern than the older streaming path.
Two interfaces, distinct expectations
Sarvam’s August 18, 2026 changelog introduced saaras:v3-realtime for voice agents and live transcription. It describes partial interim transcripts, midstream configuration and voice activity detection tuning. The provider separately maintains a legacy streaming path using saaras:v3. Reusing connection messages from one endpoint on the other can fail without any model-quality issue. Record the full endpoint contract, model ID and SDK version before implementation.
Partial text is not a final transcript
During live speech, a provisional phrase can change when the next word arrives. A caption renderer can display partial updates as an aid, but a database of legal statements, commands or meeting minutes should commit only final transcript events. Keep a monotonically increasing event sequence, tolerate reconnects and avoid appending all revisions as if they were new speech. Use the same time base for original recording and caption export.
Endpointing is a product choice
Aggressive voice activity detection reduces the sense of delay but can cut off speakers who hesitate. Conservative endpointing protects speech completeness at the cost of slower responses. Test with phone codecs, crowded environments, stutters and long pauses. For a voice assistant, an incorrect end-of-turn can be more disruptive than a slightly delayed but complete phrase; compare false interruptions separately from word error rate.
Production architecture
Use a transport adapter to normalize incoming audio frames, verify sample format and map provider events into an internal state machine. Define timeouts, cancellation and backpressure. Preserve a final transcript with chosen model and code version; do not store unbounded buffers of sensitive microphone input. Apply a human handoff when recognition fails twice, in emergencies or before irreversible account transactions.
What not to claim
A first-party changelog establishes release and interface distinctions, not objective speech accuracy or a guaranteed millisecond latency. Language and regional coverage need the exact endpoint documentation; billing may count connected time rather than successfully recognized speech. Validate all of those separately in AI10 and AI07 handoffs before an agent deployment.
Publication scope, rights and cost caveats
Research cutoff: 25 October 2026. The material above distinguishes provider documentation from independently measured behavior; no hands-on evaluation, legal clearance, voice-owner consent, commercial license purchase or model benchmark is claimed. Production teams must verify the specific API or installed version, account entitlement, input/output rights and destination market before release. Monetary plan amounts and metering tiers are not asserted here; the canonical pricing ledger belongs to AI 10. Missing release dates, exact rates and sample-rate capabilities are intentionally not inferred.