PredGen: Accelerated Inference of Large Language Models through Input-Time Speculation for Real-Time Speech Interaction

• URL: http://arxiv.org/abs/2506.15556v1
• Date: Wed, 18 Jun 2025 15:29:02 GMT
• Title: PredGen: Accelerated Inference of Large Language Models through Input-Time Speculation for Real-Time Speech Interaction
• Authors: Shufan Li, Aditya Grover,
• Abstract summary: Large Language Models (LLMs) are widely used in real-time voice chat applications, typically in combination with text-to-speech (TTS) systems to generate audio responses.<n>Their large size often leads to noticeable latency between the end of user input and the start of audio output, resulting in suboptimal user experiences.<n>We propose Predictive Generation (PredGen), a novel framework that mitigates-or even eliminates-this delay through speculative decoding at input time.
• Score: 29.64357898080842
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large Language Models (LLMs) are widely used in real-time voice chat applications, typically in combination with text-to-speech (TTS) systems to generate audio responses. However, their large size often leads to noticeable latency between the end of user input and the start of audio output, resulting in suboptimal user experiences. This latency is particularly evident when LLMs are deployed as single-user voice assistants on consumer-grade hardware with limited computing capacity. We discovered that this latency is primarily dominated by the time it takes for the LLMs to generate the first sentence, which is required as input by the TTS systems that synthesize audio responses on a sentence-by-sentence basis. To address this bottleneck, we propose Predictive Generation (PredGen), a novel framework that mitigates-or even eliminates-this delay through speculative decoding at input time. PredGen generates candidate responses while the user is still speaking, enabling the system to begin TTS processing with minimal delay. Simulated experiments on the Lmsys and MT-Bench datasets show that the proposed method can effectively reduce the latency by around 2x across a wide range of use cases, while incurring only minimal additional computation cost at input time-computation that would otherwise go unused.

Related papers

• StreamMel: Real-Time Zero-shot Text-to-Speech via Interleaved Continuous Autoregressive Modeling [50.537794606598254]
  StreamMel is a pioneering single-stage streaming TTS framework that models continuous mel-spectrograms.<n>It enables low-latency, autoregressive synthesis while preserving high speaker similarity and naturalness.<n>It even achieves performance comparable to offline systems while supporting efficient real-time generation.
  arXiv  Detail & Related papers  (2025-06-14T16:53:39Z)
• SpeakStream: Streaming Text-to-Speech with Interleaved Data [11.131427505801062]
  We present SpeakStream, a streaming TTS system that generates audio incrementally from streaming text using a decoder-only architecture.<n>During inference, SpeakStream generates speech incrementally while absorbing streaming input text.<n>Our experiments demonstrate that SpeakStream achieves state-of-the-art latency while maintaining the quality of non-streaming TTS systems.
  arXiv  Detail & Related papers  (2025-05-25T16:11:10Z)
• VITA-Audio: Fast Interleaved Cross-Modal Token Generation for Efficient Large Speech-Language Model [70.25062476543091]
  VITA-Audio is an end-to-end large speech model with fast audio-text token generation.<n>MCTP module efficiently generates multiple audio tokens within a single model forward pass.<n>Four-stage progressive training strategy is explored to achieve model acceleration with minimal loss of speech quality.
  arXiv  Detail & Related papers  (2025-05-06T17:59:53Z)
• Predictive Speech Recognition and End-of-Utterance Detection Towards Spoken Dialog Systems [55.99999020778169]
  We study a function that can predict the forthcoming words and estimate the time remaining until the end of an utterance. We develop a cross-attention-based algorithm that incorporates both acoustic and linguistic information. Results demonstrate the proposed model's ability to predict upcoming words and estimate future EOU events up to 300ms prior to the actual EOU.
  arXiv  Detail & Related papers  (2024-09-30T06:29:58Z)
• Beyond the Turn-Based Game: Enabling Real-Time Conversations with Duplex Models [66.24055500785657]
  Traditional turn-based chat systems prevent users from verbally interacting with system while it is generating responses. To overcome these limitations, we adapt existing LLMs to listen users while generating output and provide users with instant feedback. We build a dataset consisting of alternating time slices of queries and responses as well as covering typical feedback types in instantaneous interactions.
  arXiv  Detail & Related papers  (2024-06-22T03:20:10Z)
• VALL-E R: Robust and Efficient Zero-Shot Text-to-Speech Synthesis via Monotonic Alignment [101.2489492032816]
  VALL-E R is a robust and efficient zero-shot Text-to-Speech system. This research has the potential to be applied to meaningful projects, including the creation of speech for those affected by aphasia.
  arXiv  Detail & Related papers  (2024-06-12T04:09:44Z)
• Multimodal Data and Resource Efficient Device-Directed Speech Detection with Large Foundation Models [43.155061160275196]
  We explore the possibility of making interactions with virtual assistants more natural by eliminating the need for a trigger phrase. Our goal is to determine whether a user addressed the virtual assistant based on signals obtained from the streaming audio recorded by the device microphone. We address this task by combining 1-best hypotheses and decoder signals from an automatic speech recognition system with acoustic representations from an audio encoder.
  arXiv  Detail & Related papers  (2023-12-06T17:29:03Z)
• Decoder Tuning: Efficient Language Understanding as Decoding [84.68266271483022]
  We present Decoder Tuning (DecT), which in contrast optimize task-specific decoder networks on the output side. By gradient-based optimization, DecT can be trained within several seconds and requires only one P query per sample. We conduct extensive natural language understanding experiments and show that DecT significantly outperforms state-of-the-art algorithms with a $200times$ speed-up.
  arXiv  Detail & Related papers  (2022-12-16T11:15:39Z)
• High Quality Streaming Speech Synthesis with Low, Sentence-Length-Independent Latency [3.119625275101153]
  System is composed of an autoregressive attention-based sequence-to-sequence acoustic model and the LPCNet vocoder for waveform generation. Full end-to-end system can generate almost natural quality speech, which is verified by listening tests.
  arXiv  Detail & Related papers  (2021-11-17T11:46:43Z)
• Dissecting User-Perceived Latency of On-Device E2E Speech Recognition [34.645194215436966]
  We show that factors affecting token emission latency, and endpointing behavior significantly impact on user-perceived latency (UPL) We achieve the best trade-off between latency and word error rate when performing ASR jointly with endpointing, and using the recently proposed alignment regularization.
  arXiv  Detail & Related papers  (2021-04-06T00:55:11Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.