Limit 1
The locale has no dedicated LanguagePack in the runtime registry on 15 Aug 2026, so resolve_pack returns the world pack.
বাংলা · measured India voice evidence
Bengali AI voice: runtime path, real pack lines and measured latency
Bengali resolves to Bharat Essential, Google Cloud Text-to-Speech voice bn-IN-Wavenet-A, the world pack and the cascade engine. Three fresh production-adapter runs measured a 749 ms median to the first non-empty 8 kHz PCM audio chunk on 15 Aug 2026.
Measured 15 Aug 2026 · three runs · 8 kHz PCM16 · no provider fallback
Which pack, engine and vendor voice serve Bengali?
The runtime trace starts with BaseBusiness.get_voice_pack, continues through voice_config.resolve_pack and ends at voice_config.resolve_engine. The generic integration returned none from the generic business integration, resolve_pack selected the world pack, and resolve_engine selected cascade. The TTS factory then selected Google Cloud Text-to-Speech voice bn-IN-Wavenet-A, displayed as Ria, while speech recognition remained Sarvam saaras:v3.
| Runtime check | Resolved value | Meaning |
|---|---|---|
| get_voice_pack | None from the generic business integration | A client-specific pack would override the generic registry result. |
| resolve_pack | world pack | Normal replies detect and mirror language, but fixed world-pack strings remain English. |
| voice_config | Bharat Essential · Google Cloud Text-to-Speech · bn-IN-Wavenet-A | The grade write-through pins the production TTS provider and voice identifier. |
| resolve_engine | cascade | The cascade engine uses separate STT, reasoning and TTS stages. |
| speech recognition | Sarvam saaras:v3 | The primary recognition locale is bn-IN. |
The runtime table names the actual Bengali call path rather than a language-support logo: generic integration result, resolved pack, sold grade, provider voice, engine and speech-recognition provider. A client integration can replace the generic pack, but the benchmark fixture used no client override.
What first-audio latency did Bengali measure?
Bengali measured 749, 775, 632 ms across three isolated production-adapter requests. The median was 749 ms. Each run synthesized “নমস্কার, আমি আপনাকে কীভাবে সাহায্য করতে পারি?” at 8 kHz PCM16, started the timer immediately before speak(text), stopped at the first non-empty audio chunk, verified the configured adapter had not failed over, and cancelled the iterator after that first chunk.
| Run | First audio | Audio format | Adapter result |
|---|---|---|---|
| Run 1 | 749 ms | 8 kHz PCM16 | Google Cloud Text-to-Speech; no fallback |
| Run 2 | 775 ms | 8 kHz PCM16 | Google Cloud Text-to-Speech; no fallback |
| Run 3 | 632 ms | 8 kHz PCM16 | Google Cloud Text-to-Speech; no fallback |
The latency table reports a median of 749 ms from three measurements on 15 Aug 2026. The number is a reproducible TTS-adapter read, not end-to-end call latency, a percentile, or an uptime promise. Caller speech, transcription, model generation, carrier delivery and handset playback remain outside the timer.
What register does the resolved Bengali pack use?
Bengali normal replies are governed by the world pack's detect-and-mirror instruction, not a Bengali-specific fixed register. The actual fixed filler, retry and retraction strings in the resolved pack are English.
| Pack event | Exact runtime string | Register read |
|---|---|---|
| Filler while a tool runs | One moment | The fixed world-pack filler is English rather than a localized line. |
| Retry after unclear speech | Sorry, could you say that once more? | The fixed recovery line is polite English. |
| Retract an unsupported action | Sorry — I can't actually send that from here. Let me take your number and someone will get it across to you. | The fixed safety fallback is English even when the model's normal replies mirror the caller's language. |
The sample table copies three fixed strings from the world pack: the tool filler, unclear-speech retry and unsupported-action retraction. The English strings are a disclosed localization gap rather than translated marketing samples.
What are the honest Bengali limits?
Limit 2
The world pack tells the model to detect and mirror the caller's language, but its fixed filler, retry and retraction strings remain English.
Limit 3
The first-audio benchmark measures TTS adapter latency only; it excludes caller speech recognition, model reasoning, carrier transit and handset playback.
Limit 4
The WaveNet adapter is non-streaming: Google returns the full clip before the first 200 ms PCM frame is yielded.
Which evidence supports the Bengali hub?
The fresh selected-page output did not expose a Bengali row, so the Bengali route consolidates the existing Kolkata cost post without generating vertical combinations. The existing language article remains the workflow and cost depth page; the multilingual category index provides cross-language buying context; and the product page describes the broader scheduled-language surface.
Other measured Indian-language hubs
The related hubs keep one route per language and never multiply languages by verticals. Every page publishes its own runtime trace, three dated latency reads and fixed-pack evidence.
Bengali voice questions
The answers preserve the runtime resolution, measured median, exact run values and pack limitation as standalone passages that do not depend on the tables.
Bengali resolves to Bharat Essential, the Google Cloud Text-to-Speech bn-IN-Wavenet-A voice, Sarvam saaras:v3 recognition and the cascade engine. The generic integration returned no client-specific pack, so resolve_pack selected the world pack on 15 Aug 2026.
Bengali measured 749 ms median first-audio latency across three fresh 8 kHz PCM16 adapter runs: 749, 775, 632 ms. The timer started immediately before the production adapter's speak call and stopped at the first non-empty audio chunk on 15 Aug 2026.
Bengali does not have a dedicated runtime pack on 15 Aug 2026. resolve_pack returns the world pack, which asks the model to detect and mirror language but keeps its fixed filler and recovery strings in English.