Master's Thesis

Full-duplex voice dialogue system

Final Thesis 3.04 MB

Author of thesis: Ing. Sathvik Udupa

Acad. year: 2025/2026

Supervisor: Ing. Petr Schwarz, Ph.D.

Reviewer: Ing. Karel Veselý, Ph.D.

Abstract:

Cascaded spoken dialogue systems are bottlenecked by sequential
processing latency and unreliable generation on domain-specific
queries. This thesis addresses both limitations within the Unmute
full-duplex framework.

We propose Endpoint Anticipation (EPA), a fixed-horizon binary
classification task that forecasts turn completion from acoustic
features without dependence on ASR output. A multi-horizon model
(EPA-M) is integrated into Unmute via speculative execution, reducing
average response latency from 1195\,ms to 808\,ms at an Expected
Redundant Computation of 28.4\%. Evaluation against a Voice Activity
Projection baseline demonstrates consistent improvements across four
proposed metrics quantifying the trade-off between realised latency
savings and wasted speculative computation.

We further propose a latency-aware RAG integration combining
speculative retrieval with context-aware chunking, improving
retrieval hit rates by up to 54 percentage points on structured
categories. The combined system reduces end-to-end response latency
from 1520\,ms to 770\,ms relative to a retrieval-augmented baseline
while maintaining answer correctness.

Together, the contributions demonstrate that cascaded architectures
can support complex real-time processing without sacrificing
modularity or controllability.

Keywords:

full-duplex speech model, spoken dialogue system, turn-taking, retrieval-augmented-generation, multimodal interaction models

Date of defence

25.06.2026

Result of the defence

Defended (thesis was successfully defended)

znamkaAznamka

Grading

A

Process of defence

The student first presented the results he had achieved in the thesis. The committee then reviewed the supervisor’s evaluation and the reviewer’s assessment. The student then answered questions from the reviewer and additional questions from the people present at the defense. Based on the reviewer’s assessment, supervisor’s evaluation, the presentation and the the student's answers to the asked questions, the committee has evaluated the thesis with the grade A.

Topics for thesis defence

  1. Could you illustrate the "LLM-as-judge" evaluation by showing examples of reference and hypothesis inside the LLM prompt?
  2. Could you estimate the computational overhead introduced by the speculative generation ? (relative to system without endpoint anticipation)
  3. Describe  your main contribution.

Language of thesis

English

Faculty

Department

Study programme

Master of Information Technology (MIT-EN)

Composition of Committee

prof. Ing. Adam Herout, Ph.D. (předseda)
prof. Dr. Ing. Pavel Zemčík, dr. h. c. (místopředseda)
doc. RNDr. Milan Češka, Ph.D. (člen)
Ing. David Bařina, Ph.D. (člen)
doc. Ing. Vítězslav Beran, Ph.D. (člen)
Ing. Tomáš Milet, Ph.D. (člen)

Supervisor’s report
Ing. Petr Schwarz, Ph.D.

The assignment was fulfilled, and the work was significantly expanded by the Retrieval Augmented Generation (RAG) module, which accesses information from documents through speculative processing to reduce system latency. The output has practical applications in projects carried out by the Speech@FIT research group. The work was accepted for publication at the prestigious InterSpeech conference.

Evaluation criteria Verbal classification
Information about assignment

The work was challenging. The student had to study the issues of voice dialogue systems, understand the current shortcomings of state-of-the-art systems and respond to them. The work included research and implementation. The student fulfilled the assignment and, in some areas, went beyond the assignment. The work was based on the needs of research projects being solved at the faculty.

Activity during solution, consultations, communication

The work was completed well in advance and its definitive content was duly consulted.

Publication activity, awards

The work was presented at Excel@FIT. It was also accepted for publication at the prestigious conference InterSpeech.

Work with literature

The student actively approached obtaining and using study materials, primarily articles from research conferences, journals, models, and source codes from HuggingFace.

Activity during solution, consultations, communication

The student actively and continuously consulted on the work and was always properly prepared for the consultations.

Points proposed by supervisor: 95

Grade proposed by supervisor: A

Reviewer’s report
Ing. Karel Veselý, Ph.D.

Well done! This work has a very high level, proposing to post it to the competition IT SPY.

Evaluation criteria Verbal classification Points
The extent to which the requirements of the assignment have been met

Evaluation level: assignment fulfilled, and the work contains significant extensions

Extra work, not in assignment:
- Implementation of Retrieval Augmented Generation (RAG) for feeding relevant context to LLM responses

Extent of the technical report

Evaluation level: is within the usual extent

Presentation level of the technical report

the report is well structured, it reads well and can be used as introduction to the topic area

95
Formal preparation of a technical report

the language used is professional, in some places the text could be more condensed

85
Work with literature

the manuscript is richly citing research works  highly relevant to the topic

100
Realisation output

Source code is opensourced on GitHub:

https://github.com/bloodraven66/Low_latency_Spoken_Dialogue_System_RAG

The code is documented, building on top of existing tool "Unmute". It is relatively readable, although the architecture is complex. The python server is using asyncio to achieve the full-duplex streaming.

The source code also contains training code for the endpoint anticipation and service code for building the RAG knowledge base from faculty website. The code contribution is large.

90
Usability of results
  • student created an attractively designed voicebot for providing info about our faculty (courses, research groups, people, projects, etc.), this can be directly used to promote our faculty
  • other outcome is the publication at prestigious conference ASRU 2025
  • the work was also presented in Excel@FIT 2026
The difficulty of the assignment

Evaluation level: very difficult assignment

The topic is research oriented. Outcome of the work was published in prestigious conference ASRU 2025 (Article: "Streaming Endpointer for Spoken Dialogue using Neural Audio Codecs and Label-Delayed Training", where Sathvick is the 1st author)

Topics for thesis defence:
  1. Could you estimate the computational overhead introduced by the speculative generation ? (relative to system without endpoint anticipation)
  2. Could you illustrate the "LLM-as-judge" evaluation by showing examples of reference and hypothesis inside the LLM prompt?
Points proposed by reviewer: 95

Grade proposed by reviewer: A

Responsibility: Mgr. et Mgr. Hana Odstrčilová