Choosing a text-to-speech voice by its friendly name is convenient until a model family changes its catalog. Bulbul v3 and v4 should be treated as different operational surfaces.
A specific retirement notice, not an all-product shutdown
Sarvam’s October 14, 2026 changelog schedules Bulbul v3 end-of-life for Voice Agents on November 10, 2026, while explicitly retaining Bulbul v3 on the Text-to-Speech API. That means the planned November event was future-dated at the October 25 editorial cutoff. A claim that all Bulbul v3 synthesis was already retired would be inaccurate. Team owners must label which interface each application uses.
Why speaker IDs need migration
Sarvam documents model-specific voice identifiers and descriptive v4 speaker names. Sending an old v3 speaker choice into v4 may not identify the desired voice, even where the product marketing calls both “Bulbul”. Keep a source-of-truth table from legacy voice to new persona ID. For a multilingual helpdesk, make a separate expected voice per language and test naming consistently in synthesized introductions, repeated names and code-switched phrases.
Editor-controlled prosody
Prepare text with stable punctuation, number formatting and a pronunciation glossary for product names. Changing speech pace, style or pitch can alter intelligibility and how a support agent is perceived. Test in a quiet room, over typical mobile call compression and on assistive devices. Do not assume a higher model version is automatically preferred by every listener. Auditors should listen to identical scripts in randomized order.
Safe rollout before the announced EOL
Create a parallel production route for v4, preserve a configuration rollback, collect representative recorded issues and inspect interruptions. Keep voice synthesis separate from the conversational decision engine, which must not issue false reassurances during failures. Tell callers when they are speaking to an automated system where required. A voice change may itself be a brand and accessibility change, so document communications and consent reviews.
Availability and cost
Catalog entries can change by region, tenant, API version and account. The announcement establishes the Voice Agent migration date, not a permanent entitlement or billing tier. Refer numeric text-character and streaming charges to the shared AI10 snapshot. If v3 remains required for a standalone TTS service, recheck the model’s current API support before shipping rather than treating the agents notice as a universal deadline.
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.