Strongly Certified · Streaming Workflow

Multilingual Voice Support

Detects language each turn. Replies in the same language. No re-routing.

Real-time voice support that handles language switches mid-conversation. Per-turn detection, multilingual LLM and TTS, single WebSocket session. Built on the streaming-language-detect + streaming-translation contract.

Talk to an Engineer See What It Does

Per-turn

Language detection

Same-session

Mid-conversation switches

Multi-locale

STT, LLM, TTS routing

What it does

The voice loop, end-to-end.

No black box. Each step is a typed-frame node you can edit, monitor, and replace.

Caller speaks. The streaming-language-detect node tags each turn with a language code.

Conversation memory ingests the turn with its language. The prompt template injects locale-aware instructions.

The multilingual LLM answers in the detected language. The matching TTS voice synthesises the reply.

Next turn, the language can change. The router handles it. The session never re-routes.

Capabilities

Built for production. Day Two-ready.

Streaming graph contract, observability, and cost discipline come standard. The agent ships with a full test suite that runs in CI on every node version bump.

Per-turn detection

streaming-language-detect tags every transcribed turn with an ISO language code and a confidence. Below-threshold turns fall back to the session default - no silent mis-routing.

ISO codesConfidence-gatedPer-turn

Mid-call language switch

The same WebSocket session handles English → Spanish → French without dropping the call. Conversation memory keeps a unified history; only the current turn's locale changes.

Single sessionUnified memoryNo re-routing

Locale-aware prompt

The streaming-prompt-template injects per-language instructions into the system prompt. Tone, register, formality - locale-specific defaults you can edit in one place.

Template-drivenEditableLocale defaults

Streaming voice loop

STT, language-detect, prompt, LLM, and TTS run concurrently behind the typed-frame contract. Latency stays under the bilingual budget.

ConcurrentTyped framesADR-S11

Per-language model swap

The install wizard exposes the LLM and TTS as separate fields per language. Use a multilingual frontier model for everything, or pin a smaller model for high-volume locales.

Per-locale configWizard-drivenCost flex

Live span tree

Each turn's spans carry the detected language as an attribute. Filter the canvas overlay by locale to debug accent issues, regional vocab, or low-confidence detections.

ADR-S14Lang attributeCanvas filter

Built on

Real services. Your stack.

Every dependency is a registered Strongly service or a model you control. Swap any one of them in the install wizard. The graph stays intact.

STT model

Multilingual STT (whisper-1 multilingual default)

Language detect

streaming-language-detect node - confidence-gated

LLM model

Multilingual chat model (gpt-4o multilingual default)

TTS model

Locale-aware TTS routing (tts-1 multi-voice)

Five common customisations

Tune it. Don't fork it.

The marketplace template is the graph. Every customisation below is a config change or a single-node addition - never a rewrite.

Per-language LLM routing

Replace the single llm node with a streaming-conditional + branch-per-locale. Cheap models for high-volume languages, premium models for the long tail.

Custom locale list

The default supports the multilingual model's full set. Restrict to your supported locales in the language-detect node's allow-list.

Translation fallback

If a TTS voice for the detected language isn't available, chain streaming-translation to render the reply into a supported language and disclose the translation.

Domain glossary

Inject a locale-aware glossary (technical terms, brand names, do-not-translate list) into the prompt template per turn.

Compliance auditing

Add streaming-conversation-store tagged with the per-turn language for regulator-friendly transcripts.