Sign InOpen Brain
arXivPaperNeeds Review

Spoken Function Calling: A New Perspective on Spoken Language Understanding for Large Audio Language Models

Spoken Function Calling reframes speech understanding as structured function selection, giving voice-agent builders a clearer interface for open-domain tool invocation.

arXiv · Aug 5, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Spoken Function Calling replaces ambiguous intent rules with structured function definitions. The work curates spoken functions from existing datasets, uses a **multi-agent system** to synthesize **SFC-Bench**, evaluates text and audio language models, and post-trains audio models for the task.

Practical Implication

Voice-agent builders should consider representing speech semantics directly as callable functions instead of maintaining a separate closed-set intent layer. That structure may make in-context examples and tool contracts easier to extend across domains.

Agent-Ready Context
Spoken Function Calling replaces ambiguous intent rules with structured function definitions. The work curates spoken functions from existing datasets, uses a **multi-agent system** to synthesize **SFC-Bench**, evaluates text and audio language models, and post-trains audio models for the task.

Voice-agent builders should consider representing speech semantics directly as callable functions instead of maintaining a separate closed-set intent layer. That structure may make in-context examples and tool contracts easier to extend across domains.

The abstract says SFC improves semantic extraction over traditional spoken-language understanding but provides no dataset size, accuracy figures, or error breakdown. Its robustness to real microphones, accents, noise, and production tool schemas is therefore unresolved.
Connected Context · Feed7 Judgment

Spoken Function Calling moves voice understanding from fixed intent classification toward extensible tool contracts, making structured calls part of semantic extraction rather than a downstream translation step. It complements realtime voice infrastructure and structured device tools, while leaving production reliability under noise, accents, and real schemas unevaluated.

AI Gateway now supports streaming transcriptionStreaming transcription offers a text-mediated path from speech into existing agents; Spoken Function Calling instead trains and evaluates models to map spoken input directly onto structured function definitions.Grok Voice Think Fast 2.0 now available on AI GatewayGrok Voice provides realtime speech-to-speech execution and early tool calls, while SFC supplies a benchmark and post-training formulation for assessing whether spoken semantics are converted into the correct structured call.PalmClaw: A Native On-Device Agent Framework for Mobile PhonesPalmClaw shows the execution side of exposing capabilities as structured calls; SFC addresses the preceding interpretation problem of selecting and populating such calls directly from speech.Stop AI Agent Hallucinations: 5 Techniques + Production Patterns - Elizabeth Fuentes, AWSStructured function definitions create the narrow tool contracts recommended by the reliability controls, but SFC still needs validation and pre-call rules because improved extraction does not establish safe execution.
Context Map
agentaudio#tool-use#agent-evals
Uncertainty
The abstract says SFC improves semantic extraction over traditional spoken-language understanding but provides no dataset size, accuracy figures, or error breakdown. Its robustness to real microphones, accents, noise, and production tool schemas is therefore unresolved.