/ai

Routing voice AI tool calls through a text backend

A post discussing NVIDIA research reports a gap on grounded customer-service tasks under clean conditions: commercial duplex voice models complete 31–51%, versus 85% for text agents such as GPT-5 on the same tasks in text mode.

—

EL

1 Source, 11h ago, first seen 11h ago

TLDR

A post describes an NVIDIA research architecture that delegates tool decisions to a text model rather than the speech model. The full-duplex speech frontend streams transcripts to the text backend, which handles the tool call; the result is passed back for text-to-speech output. The post reports 92.0–97.2% tool-call recall and 81.2% accuracy at rejecting irrelevant calls. It says turn-taking rate, streaming speech-recognition word error rate and spoken-language intelligence remain at the no-tool-call baseline.

Today's Rank

—

Not ranked yet

Combined views

7K

1 Source, first seen 11h ago

1 Source

@omarsar0NVIDIA research papers are on fire recently! Here is another interesting paper where they give full-duplex speech models tool calls. (bookmark it) Commercial duplex voice models complete 31 to 51 percent of grounded customer-service tasks under clean conditions. Text agents like GPT-5 reach 85 percent on the same tasks in text mode. Most of what a voice agent loses, it loses in the speech pipeline. The fix routes the decision out of the speech model. The duplex frontend learns to emit a delegation token, forwards streaming transcripts to a text backend LLM for the tool call, and receives the result through a lightweight prefill-and-repeat mechanism before streaming TTS speaks it. Tool-call recall runs 92.0 to 97.2 percent with 81.2 percent accuracy at rejecting irrelevant calls. Turn-taking rate, streaming ASR word error rate and spoken-language intelligence all stay at the no-tool-call baseline. Paper: https://arxiv.org/abs/2609.19334 Chat with Paper: https://academy.dair.ai/papers/a-frontend-backend-architecture-for-tool-calls-in-full-duplex-speech-models-2609.19334

Today's Rank

—

Not ranked yet

Combined views

7K

1 Source, first seen 11h ago

68 likes

19 comments

69 saves

10 reposts

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

TLDR

A post describes an NVIDIA research architecture that delegates tool decisions to a text model rather than the speech model. The full-duplex speech frontend streams transcripts to the text backend, which handles the tool call; the result is passed back for text-to-speech output. The post reports 92.0–97.2% tool-call recall and 81.2% accuracy at rejecting irrelevant calls. It says turn-taking rate, streaming speech-recognition word error rate and spoken-language intelligence remain at the no-tool-call baseline.

19 comments

69 saves

10 reposts

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

1 Source

@omarsar0NVIDIA research papers are on fire recently! Here is another interesting paper where they give full-duplex speech models tool calls. (bookmark it) Commercial duplex voice models complete 31 to 51 percent of grounded customer-service tasks under clean conditions. Text agents like GPT-5 reach 85 percent on the same tasks in text mode. Most of what a voice agent loses, it loses in the speech pipeline. The fix routes the decision out of the speech model. The duplex frontend learns to emit a delegation token, forwards streaming transcripts to a text backend LLM for the tool call, and receives the result through a lightweight prefill-and-repeat mechanism before streaming TTS speaks it. Tool-call recall runs 92.0 to 97.2 percent with 81.2 percent accuracy at rejecting irrelevant calls. Turn-taking rate, streaming ASR word error rate and spoken-language intelligence all stay at the no-tool-call baseline. Paper: https://arxiv.org/abs/2609.19334 Chat with Paper: https://academy.dair.ai/papers/a-frontend-backend-architecture-for-tool-calls-in-full-duplex-speech-models-2609.19334

Today's Rank

—

Not ranked yet

Today's Rank

—

Not ranked yet