AI assistants on Larsa AI can speak in two distinct modes.
Each mode affects how a caller’s speech is understood and how the assistant’s reply is generated:

1. Pipeline

Label in UI: Pipeline
How it works: Speech-to-Text → LLM → Text-to-Speech
Latency: ~800 – 1500 ms (varies by language & model)
Best for: Complex reasoning, dynamic prompts, multi-sentence replies

Overview:
Pipeline mode first transcribes the caller’s words into text, processes them with the language model, then converts the response back into audio. This traditional approach offers maximum flexibility:

• Works with all voices in the library (including custom-cloned voices).

• Great for long-form, detailed, or paragraph-style answers.

• Easily references earlier context and injects variables.

When to choose Pipeline:

• You need rich, multi-sentence answers (e.g., support queries, detailed explanations).

• The assistant must reason over structured data or complex prompts.

• You want complete control over the spoken voice (clone or brand voice).

2. Speech-to-Speech (Multimodal)

Label in UI: Speech-to-Speech
How it works: Direct speech-to-speech generation (no text stage)
Latency: ~300 – 600 ms (ultra-low)
Best for: Natural back-and-forth, short and reactive replies

Overview:
Speech-to-Speech skips separate transcription and TTS, using a multimodal model that listens and responds directly for smoother conversations:

• Near-instant responses with fast turn-taking.

• Naturally expressive speech (intonation, pauses, fillers).

• Limited voice options for now, but expanding regularly.

When to choose Speech-to-Speech:

• You want snappy, conversational flow (sales, booking confirmations).

• Responses are typically short sentences or quick acknowledgments.

• You’re fine using the system-provided voices for faster interaction.

Note: If you need a custom cloned voice or advanced prompt logic, Pipeline is still the better choice.

Switching Modes

You can select the mode for each assistant in Assistant → Settings → Voice Engine.
Test both modes to find the best balance of speed and quality for your use case.

Pro Tip: Record two calls—one in each mode—and compare perceived latency and engagement to decide which works best.