An End-to-End Generative System for Smart Travel Assistant
Mirac¸ Tu
˘
gcu
1
, Beg
¨
um C¸ ıtamak Erdinc¸
1
, Tolga C¸ ekic¸
1
, Seher Can Akay
1
, Derya Uysal
1
, Onur Deniz
1
and Erkut Erdem
2
1
Natural Language Processing Department, Yapı Kredi Teknoloji, Istanbul, Turkey
2
Department of Computer Engineering, Hacettepe University, Ankara, Turkey
Keywords:
Generative AI, Voice Assistant, Text-to-Speech, Speech-to-Text, Chatbot, Language Models, Deep Learning,
Natural Language Processing.
Abstract:
Planning a travel with a customer assistant is a multi-stage process that involves information collecting, and
usage of search and reservation services. In this paper, we present an end-to-end system of a voice-enabled
virtual assistant specifically designed for travel planning in Turkish. This system involves fine-tuned state-of-
the-art models of Speech-to-text (STT) and Text-to-speech (TTS) models for increased success in the tourism
domain for Turkish language as well as improvements to chatbot experience that can handle complex, mul-
tifaceted conversations that are required for planning a travel thoroughly. We detail the architecture of our
voice-based chatbot, focusing on integrating STT and TTS engines with a Natural Language Understanding
(NLU) module tailored for travel domain queries. Furthermore, we present a comparative evaluation of speech
modules, considering factors such as parameter size and accuracy. Our findings demonstrate the feasibility of
voice-based interfaces for streamlining travel planning and booking processes in Turkish language which lacks
high-quality corpora of speech and text pairs.
1 INTRODUCTION
When planning their trips, users encounter a range of
options and constraints. Traditionally, the planning
process relies mostly on online search engines and
user interactions via an interface which can be cum-
bersome. Voice assistants offer a more flexible and
accessible way for users to express their needs. Im-
proving human-computer interaction is possible by
developing such an interface with a virtual assistant
to offer a natural and intuitive way to provide infor-
mation. A voice-enabled assistant has the potential
to significantly improve this experience by allowing
users to verbally convey their needs, and receive both
textual and spoken confirmations about booking de-
tails.
The main objective of such an assistant system is
understanding the requests of user and perform an ac-
tion related to a travel topic. Therefore, intent clas-
sification and slot-filling, which are two crucial NLU
components, are used to decide which travel related
function to perform e.g. searching for tours, booking
a hotel, cancelling reservations and so on. In (D
¨
undar
et al., 2020), a robust intent classifier for Turkish lan-
guage is proposed with a similar objective for the
banking domain. However, a slot-filling module is
also needed to perform an action related to intention
based on the preferences of a user. To meet this need,
a named entity recognition (NER) model can be in-
tegrated into the chatbot. A recent work (Stepanov
and Shtopko, 2024) demonstrates a specialized trans-
former model that outperforms ChatGPT and fine-
tuned LLMs in zero-shot cross-domain NER bench-
marks for various languages except Turkish. Users
might specify their preferences in a more natural man-
ner where contextual relation and domain knowledge
are required. With this purpose, slot-filling can be
even more successful in a few-shot setting with LLMs
(Brown et al., 2020) instead of zero-shot.
To understand the user’s intention, we uti-
lized a BERT (Bidirectional Encoder Representations
from Transformers) classifier (Devlin et al., 2019),
which has been specifically fine-tuned for Turkish
(Schweter, 2020). BERT is well-suited for under-
standing context and nuance in a language due to its
deep bidirectional architecture. This allows the model
to consider the full context of a word by looking at
words that come before and after it. This is partic-
Tu
ˇ
gcu, M., Erdinç, B., Çekiç, T., Akay, S., Uysal, D., Deniz, O. and Erdem, E.
An End-to-End Generative System for Smart Travel Assistant.
DOI: 10.5220/0013064200003838
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 16th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2024) - Volume 1: KDIR, pages 473-479
ISBN: 978-989-758-716-0; ISSN: 2184-3228
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.
473

ularly beneficial for agglutinative languages such as
Turkish.
There are notable challenges to developing a
voice-enabled travel assistant in Turkish, due to
the lack of natural voice generation and Automatic
Speech Recognition (ASR) models which are also ro-
bust to noise and low-quality voice sources. This is
mostly because of the limited availability of high-
quality parallel data for training robust speech recog-
nition and synthesis models in Turkish. There are
multi-lingual TTS models successful in generating
natural speech or speech recognition e.g. XTTS
(Casanova et al., 2024) and Whisper (Radford et al.,
2023), however, the models either have licences not
available for commercial use or demand high com-
putational resources. The latency of a response gen-
erated by a smart assistant directly affects the user
experience. Therefore, a mono-lingual and single
speaker but a robust, small architecture satisfies the
high-throughput need such as MMS (Pratap et al.,
2024) and FastSpeech2 (Ren et al., 2020). For au-
tomatic speech recognition task, there are successful
models introduced in recent years such as Wav2Vec
2.0 (Baevski et al., 2020) and Whisper (Radford et al.,
2022) with multi-lingual foundation models available.
However, this foundation models has only rudimen-
tary capabilities in some languages such as Turkish
and they require further fine tuning to perform well
enough for active usage.
In this study, a chatbot with NLU modules such
as intent classification and slot filling in the travel do-
main for searching, booking and purchasing purposes
of hotels and tours is developed. We approached
the slot-filling problem with a hybrid approach by
using few-shot prompting technique with an LLM
where context matters the most for user messages. We
further trained robust and lightweight STT and TTS
models for Turkish language in the tourism domain to
develop a voice interface for the chatbot which com-
pletes the virtual assistant experience.
2 SYSTEM ARCHITECTURE
Visual representation of our developed system is il-
lustrated at Figure 1. The user is able communicate
with the assistant through speech modules or writ-
ten chat. Conversation flow manager is a multi mod-
ule system that understands the intention of the user,
leverages generative slot-filling and pattern matching
to perform an action with given information through
travel services. Through the function calling com-
ponent located in the conversation flow manager, in-
tent classification, slot filling, and pattern matching
User Message
Response
User
Speech Speech
STT
Post Correction
Voice
Conversion
TTS
Intent
Classification
Generative
Slot Filling
Function
Calling
Pattern
Matching
Conversation Flow Manager
LLM
Travel Services
Figure 1: Overall virtual assistant architecture.
components elucidated in Section 2.1 enable the se-
mantic interpretation of transcribed sentences. The
generative slot-filling and pattern matching compo-
nents address different needs by employing distinct
methodologies. Generative slot filling leverages gen-
erative models to identify entities that cannot be eas-
ily expressed through predefined rules, whereas the
pattern matching component uses regular expressions
and fuzzy match scores based on predefined dictio-
naries to detect entities with fixed formats, such as
hotel names, cities, districts, and dates. By lever-
aging the information extracted from these sentences
within travel planning services, the system facilitates
user interaction, ensuring the effective execution of
functions such as system utilization and information
retrieval from services. Moreover, with the function
calling component, we enabled dynamic modification
of endpoints and service variables directly from the
interface we designed. This approach allowed for the
seamless integration of new services with intents and
facilitated the rapid adaptation to changing service re-
quirements without the need for additional coding.
2.1 NLU Module
The Natural Language Understanding (NLU) compo-
nent of our chatbot has two major components: intent
classification and entity recognition. For intent clas-
sification, we utilized the BERT model based on the
KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval
474

methodology outlined in (D
¨
undar et al., 2020). We
employed BERTurk model (Schweter, 2020), which
is a model trained in Turkish corpora. This BERT
based classifier met our demands and surpassed few
shot training with LLMs, hence it is finetuned to be
used as an intent classifier in travelling domain for
this work as well.
On the other hand, we have observed that the en-
tity collection for tourism can be challenging. The
words we consider as entities can vary significantly
in terms of subject matter and type. It may be
necessary to perform entity extraction for a diverse
range of entities, such as spa, sport, aquapark,
nature, outdoor pool, child/baby-friendly,
pet-friendly, honeymoon, and seafront. A user
may wish to specify multiple features of the desired
hotel within a single sentence to obtain results based
on those criteria. Since it is more appropriate to
consider these words as features rather than distinct
entities, they were tagged as feature1, feature2
and so forth, before being transmitted to the rele-
vant services. Additionally, the sentences constructed
by users do not adhere to specific rhetorical patterns.
Due to these problems, we decided that the use of
large language models is more appropriate for this
problem because of their capability of understanding
complex patterns with fewer training examples.
To achieve this, we utilize ChatGPT from Ope-
nAI to extract entities from sentences with our genera-
tive slot-filling component. By engineering a dynamic
prompt, we were able to receive a JSON-formatted
output that parsed the specified types and numbers
of entities from the given sentences. Additionally,
through the modification capabilities provided within
the application, we enabled the addition or removal of
new entities without the need for further development.
Moreover, working with large language models
inherently posed the risk of receiving outputs in irreg-
ular formats. To mitigate this, we provided JSON ex-
amples within the prompts and implemented checks
to ensure the outputs adhered to JSON rules. Ad-
ditionally, since user prompts were directly fed into
the large language model for entity extraction, this
opened the possibility for the system’s outputs to be
manipulated. To address this, we refined the prompts
to prevent users from altering the system prompt and
obtaining distorted results.
Additionally, as mentioned in Section 2, we en-
sured that entities, which could be defined by rules,
were identified for travel services by comparing them
against regular expressions and words in our cus-
tom dictionaries, using calculated fuzzy match scores.
The system we developed is depicted in Figure 2.
User
"I want to book a hotel
with a spa in Bodrum,
surrounded by trees"
Conversation
Flow Manager
Generative
Slot Filling
Pattern
Matching
OpenAI Services
Travel Services
[{'text': 'spa', 'type':
'spa'}, {'text': '{trees},
'type': 'nature'}]
["district":"Bodrum"]
Figure 2: Extracting entities from user prompts.
2.2 Text-to-Speech Module
The presented generation pipeline, as shown in Fig-
ure 3, contains a phoneme encoder, the LightSpeech
TTS model, a vocoder and a voice conversion model.
The text is converted to phoneme sequence by using
an open-source Turkish grapheme-to-phoneme model
and dictionary (McAuliffe et al., 2017). Speech syn-
thesis with low latency is necessary for a seamless
user experience which is why we use LightSpeech
(Luo et al., 2021) and Parallel WaveGAN (PWG)
(Yamamoto et al., 2020) vocoder that proved its effi-
ciency. LightSpeech model is based on FastSpeech 2
but its architecture is designed more lightweight and
more efficient via Neural Architecture Search. The
audio quality is on par with FastSpeech2 while having
a remarkable inference speed up. The generated mel-
spectrograms are transformed into audio waveform by
using a Parallel WaveGAN vocoder that is pre-trained
on LibriTTS (Zen et al., 2019) which is capable of
high-fidelity speech generation for Turkish. Finally,
the OpenVoice (Qin et al., 2023) model is utilized for
zero-shot cross-lingual Voice Cloning to specifically
convert the speaker of the generated audio waveform.
This pipeline allows alternative models to be used in
An End-to-End Generative System for Smart Travel Assistant
475

Chatbot
LightSpeech
Phonemizer
Voice Conversion
Vocoder
Target Voice
User
Speech
Phonemes
Graphemes
Converted Speech
Figure 3: Text-to-speech pipeline for generation and voice
cloning.
any part. For example, a vocoder model or a differ-
ent voice cloning model can be easily implemented to
replace respective parts.
2.3 Speech-to-Text Module
For the speech-to-text module Wav2Vec 2.0 speech
recognition model is used and in order to correct po-
tential errors in transcripts a post-correction method
using an N-gram language model is used as shown in
Figure 4. Wav2Vec 2.0 is a transformer based model
that can be trained with raw audio data without any
need for preprocessing (Baevski et al., 2020). Us-
ing raw audio data helps both with managing training
data and with inference in the software pipeline as it
does not introduce another layer that increases com-
plexity. Using a multi-lingual foundation model with
this architecture we have fine-tuned the model Turk-
ish data and tourism related data. We have also im-
plemented another layer for post-correction using N-
gram based language model KenLM (Heafield, 2011).
KenLM is a fast language modelling tool that can
be used to create N-gram language models efficiently
and also can be adapted to work with the Wav2Vec
2.0 model. Post-correction layer is used not only be-
cause it helps with correcting transcription errors that
may arise due to similar sounding words and exter-
nal noises; but also, recent studies have shown that
using N-gram language model based post-correction
can improve performance in low resource languages
User
Wav2Vec2
m e r a b a
KenLM
m e r h a b a
Chatbot
Figure 4: Speech-to-text pipeline that contains sequence
modelling and post-correction models. As a side note,
merhaba means hello in Turkish.
(Avram et al., 2023) and it can help with better adap-
tation on specific domains (Ma et al., 2023). We have
created 5-gram language models to use in our experi-
ments from general domain and tourism domain texts.
The other model we experimented on is W2v-
BERT which combines the language model post-
correction aspect into the trained model (Chung et al.,
2021). This model uses a BERT encoder model as
a language model instead of an N-gram language
model. The advantage of using BERT is that it keeps
a larger contextual information and also semantic
knowledge of the words as well but compared to us-
ing an N-gram model it is a more resource demanding
approach.
3 EXPERIMENTS & RESULTS
3.1 Text-to-Speech Experiments
Experimental Setup. We evaluate the LightSpeech
model trained on a dataset that contains 5,131 audio
samples with approximately 6 hours of novel reading
without their text pairs. The average duration of the
audio samples is 4.1 seconds. The transcriptions of
audio samples are generated using our STT method.
The errors in synthetic data consist mostly of similar-
sounding words. Therefore, the effects of these errors
are very little. Moreover, the errors in the synthetic
transcriptions are expected to be minimal due to audio
quality. The speech dataset and its phonemes are gen-
erated and aligned with the Montreal Forced Aligner
public tool (McAuliffe et al., 2017) following (Ren
KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval
476

et al., 2020). The audio waveforms are transformed
into mel-spectrograms following (Luo et al., 2021),
differently, we set the frame size and hop size to 300
and 1200 concerning the sampling rate of 24000. We
train the model for 100k steps on a single NVIDIA
V100 GPU. Models in our TTS pipeline other than the
LightSpeech model are utilized as pre-trained models
with their public weights.
Evaluation Methodology. There is no straightfor-
ward approach to evaluate speech generation. Most
of the speech features like timbre or prosody may
vary in the generated speech of a text compared to
the ground truth utterance and it is an even harder
challenge for multi-speaker datasets. Therefore, it
is meaningful to evaluate a system by the aspect or
the feature needed. We decided to evaluate intelli-
gibility and pronunciation by transcribing the gen-
erated speech with ASR. We specifically choose a
well-known and capable multi-lingual model Whisper
(Radford et al., 2023), and 743 audio-text pairs from
the Turkish subset of multi-lingual ASR benchmark
known as FLEURS (Conneau et al., 2023) which is
considered as an out-of-domain evaluation with re-
spect to our training domain. The subset is approx-
imately 2.6 hours long and the average duration of
samples is 12.6 seconds. We generate speech of the
texts from the dataset to create synthetic audio and
original text pairs for each TTS model. ASR models
transcribe TTS outputs into text hypotheses, allow-
ing us to calculate the Word Error Rate (WER) and
Character Error Rate (CER) by comparing them to the
original transcript. For measuring the error rates, we
apply the normalization of Whisper on references and
hypotheses. In most cases, another preferable evalu-
ation method is to evaluate naturalness and audio fi-
delity by Mean Opinion Score (MOS) metric but it’s
not an automatic evaluation strategy and the reliance
on human raters presents a challenge. However, we
decided to use a subset of 100 utterances generated for
each model from the ASR benchmark we mentioned.
We compare our results with public models success-
ful in Turkish speech synthesis: 1) pre-trained Turk-
ish MMS TTS model (Pratap et al., 2024) which is an
end-to-end model with VITS (Kim et al., 2021) archi-
tecture, and 2) multi-lingual XTTS (Casanova et al.,
2024) model with zero-shot voice-cloning feature that
has a novel architecture based on Tortoise (Betker,
2023) and a HiFi-GAN vocoder (Kong et al., 2020)
with 26M parameters. Parameter sizes of models are
shown in Table 1.
Results. In our experiments, the LightSpeech
model (our setup) is able to generate utterances that
Table 1: Text-to-speech models that is used in experiments
and their parameter size.
Model #Params
LightSpeech 1.8M
LightSpeech + PWG 3.1M
MMS 36.3M
XTTS 466.9M
Table 2: Evaluation results of Turkish speech synthesis by
using Whisper models and FLEURS benchmark dataset.
Original denotes the results of ASR models from Whisper
paper (Radford et al., 2023).
Model WER(↓) CER(↓)
Whisper-medium
LightSpeech 13.0 2.9
MMS 18.4 4.4
XTTS 10.1 2.5
Original 10.1 -
Whisper-large-v2
LightSpeech 10.8 2.5
MMS 15.3 3.7
XTTS 8.3 2.5
Original 8.4 -
Table 3: MOS scores from a human study with regard to
naturalness on a subset of Turkish FLEURS dataset.
Model MOS(↑)
LightSpeech 2.98
±0.081
MMS 3.34
±0.082
XTTS 4.43
±0.055
preserve the text content better than the MMS TTS
model, as shown in Table 2. LightSpeech performs
less well than XTTS which has equal or better ac-
curacy on ASR evaluation than the original utter-
ances in the FLEURS dataset. However, our setup
is as accurate as XTTS on the CER metric evaluation
with Whisper-large-v2. Also, there is a slight differ-
ence with XTTS on the CER metric evaluation with
Whisper-medium. However, the MOS score of Light-
Speech is less natural than MMS and far from XTTS
on naturalness as shown in Table 3. This is mostly
due to the size of the training data and its record-
ing quality. Also, our observations show that our
model is less natural on long input sequences of the
FLEURS benchmark because it is trained on a dataset
with short sequences and is not able to generalize long
sequences in terms of naturalness. Note that MMS
and XTTS models have nearly 12x and 150x more
parameters than LightSpeech + PWG respectively, as
shown in Table 1. Therefore, the results show that the
model is robust in comprehensibility but needs im-
provement on naturalness, considering the constraints
imposed by its size and limited training data.
An End-to-End Generative System for Smart Travel Assistant
477

Table 4: Performance Comparison of Speech Recognition
Models.
General Test Set
Model WER(%) CER(%)
Wav2Vec 2.0 14.038 4.070
W2v-BERT 16.636 4.252
Wav2Vec 2.0 + kenLM 8.106 1.669
Tourism Test Set
Model WER(%) CER(%)
W2v-BERT 13.112 2.229
Wav2Vec 2.0+kenLM 8.888 1.974
3.2 Speech-to-Text Experiments
Experimental Setup. The evaluation was done on a
6 hours dataset that is obtained from the public Turk-
ish Common Voice dataset in the general domain and
1 hours dataset from tourism domain. For evalua-
tion of Speech-to-text tasks, generally used metrics
are word error rate (WER) and character error (CER).
These metrics measure the error rates of transcriptions
compared to the actual transcriptions of audio files.
The lower error rates mean the model is more suc-
cessful.
Results. Our experiments have demonstrated that
using Wav2Vec 2.0 model together with kenLM post-
correction outperforms using it without the language
model and W2v-BERT model. The results are shown
in shown in table 4. It is unsurprising for the N-gram
language model post-correction to surpass the perfor-
mance of the base model as the previous studies have
shown similar results. The low scores from W2v-
BERT model may be due to the multi-lingual foun-
dation model’s BERT model not being too successful
in Turkish language.
4 CONCLUSION & FUTURE
WORK
In this paper, we introduced the pipeline for a voice
assistant in Turkish, that is capable of helping users
in the tourism domain. This assistant leverages an in-
tuitive voice interface by enabling users to seamlessly
request information, access travel services, and com-
plete their entire travel planning experience through
spoken interactions. For slot-filling task of the as-
sistant, a hybrid approach that combines regular ex-
pressions with few-shot LLM prompting is utilized.
Additionally, lightweight and robust models for our
NLU and speech modules are implemented to en-
sure a conversation at a natural pace. Our findings
have demonstrated that the speech-to-text and text-to-
speech models we trained achieved high intelligibility
in spite of the scarcity of Turkish speech resources.
For future work, to improve the performance of
text-to-speech models we intend to increase the qual-
ity and the quantity of our training data by speech en-
hancement and denoising techniques. We also aim
to implement a zero-shot prosody cloning feature to
the TTS pipeline to control the emotion emphasized
in synthesized speech. For speech recognition, an ad-
ditional post-correction model will be used to correct
transcriptions of foreign words that can often be en-
countered in the tourism domain. For the NLU com-
ponent, which constitutes the chatbot’s understanding
functions, we aim to leverage generative methods fur-
ther to provide the user with more diverse and varied
responses.
ACKNOWLEDGEMENTS
This work was done as a part of Smart Travel
Digital Ecosystem project which was supported by
TUBITAK (no: 9220043) and Celtic-NEXT (no:
CE-2020-2-3) for the international industrial R&D
projects program.
REFERENCES
Avram, A.-M., Sm
˘
adu, R.-A., P
˘
ais¸, V., Cercel, D.-C., Ion,
R., and Tufis¸, D. (2023). Towards improving the
performance of pre-trained speech models for low-
resource languages through lateral inhibition.
Baevski, A., Zhou, H., Mohamed, A., and Auli, M. (2020).
wav2vec 2.0: A framework for self-supervised learn-
ing of speech representations.
Betker, J. (2023). Better speech synthesis through scaling.
arXiv preprint arXiv:2305.07243.
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.,
Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G.,
Askell, A., Agarwal, S., Herbert-Voss, A., Krueger,
G., Henighan, T., Child, R., Ramesh, A., Ziegler,
D. M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler,
E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner,
C., McCandlish, S., Radford, A., Sutskever, I., and
Amodei, D. (2020). Language models are few-shot
learners. CoRR, abs/2005.14165.
Casanova, E., Davis, K., G
¨
olge, E., G
¨
oknar, G., Gulea,
I., Hart, L., Aljafari, A., Meyer, J., Morais, R.,
Olayemi, S., et al. (2024). Xtts: a massively multi-
lingual zero-shot text-to-speech model. arXiv preprint
arXiv:2406.04904.
Chung, Y.-A., Zhang, Y., Han, W., Chiu, C.-C., Qin, J.,
Pang, R., and Wu, Y. (2021). W2v-bert: Combining
contrastive learning and masked language modeling
for self-supervised speech pre-training.
KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval
478

Conneau, A., Ma, M., Khanuja, S., Zhang, Y., Axelrod,
V., Dalmia, S., Riesa, J., Rivera, C., and Bapna, A.
(2023). Fleurs: Few-shot learning evaluation of uni-
versal representations of speech. In 2022 IEEE Spoken
Language Technology Workshop (SLT), pages 798–
805. IEEE.
Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K.
(2019). BERT: Pre-training of deep bidirectional
transformers for language understanding. In Pro-
ceedings of the 2019 Conference of the North Amer-
ican Chapter of the Association for Computational
Linguistics: Human Language Technologies, Volume
1 (Long and Short Papers), pages 4171–4186, Min-
neapolis, Minnesota. Association for Computational
Linguistics.
D
¨
undar, E. B., Kilic¸, O. F., Cekic¸, T., Manav, Y., and Deniz,
O. (2020). Large scale intent detection in turkish short
sentences with contextual word embeddings. In KDIR,
pages 187–192.
Heafield, K. (2011). KenLM: Faster and smaller language
model queries. In Callison-Burch, C., Koehn, P.,
Monz, C., and Zaidan, O. F., editors, Proceedings of
the Sixth Workshop on Statistical Machine Transla-
tion, pages 187–197, Edinburgh, Scotland. Associa-
tion for Computational Linguistics.
Kim, J., Kong, J., and Son, J. (2021). Conditional varia-
tional autoencoder with adversarial learning for end-
to-end text-to-speech. In International Conference on
Machine Learning, pages 5530–5540. PMLR.
Kong, J., Kim, J., and Bae, J. (2020). Hifi-gan: Genera-
tive adversarial networks for efficient and high fidelity
speech synthesis. Advances in neural information pro-
cessing systems, 33:17022–17033.
Luo, R., Tan, X., Wang, R., Qin, T., Li, J., Zhao, S., Chen,
E., and Liu, T.-Y. (2021). Lightspeech: Lightweight
and fast text to speech with neural architecture search.
In ICASSP 2021-2021 IEEE International Confer-
ence on Acoustics, Speech and Signal Processing
(ICASSP), pages 5699–5703. IEEE.
Ma, R., Wu, X., Qiu, J., Qin, Y., Xu, H., Wu, P., and Ma,
Z. (2023). Internal language model estimation based
adaptive language model fusion for domain adapta-
tion.
McAuliffe, M., Socolof, M., Mihuc, S., Wagner, M., and
Sonderegger, M. (2017). Montreal forced aligner:
Trainable text-speech alignment using kaldi. In In-
terspeech, volume 2017, pages 498–502.
Pratap, V., Tjandra, A., Shi, B., Tomasello, P., Babu, A.,
Kundu, S., Elkahky, A., Ni, Z., Vyas, A., Fazel-
Zarandi, M., et al. (2024). Scaling speech technology
to 1,000+ languages. Journal of Machine Learning
Research, 25(97):1–52.
Qin, Z., Zhao, W., Yu, X., and Sun, X. (2023). Open-
voice: Versatile instant voice cloning. arXiv preprint
arXiv:2312.01479.
Radford, A., Kim, J. W., Xu, T., Brockman, G., McLeavey,
C., and Sutskever, I. (2022). Robust speech recogni-
tion via large-scale weak supervision.
Radford, A., Kim, J. W., Xu, T., Brockman, G., McLeavey,
C., and Sutskever, I. (2023). Robust speech recogni-
tion via large-scale weak supervision. In International
conference on machine learning, pages 28492–28518.
PMLR.
Ren, Y., Hu, C., Tan, X., Qin, T., Zhao, S., Zhao, Z.,
and Liu, T.-Y. (2020). Fastspeech 2: Fast and high-
quality end-to-end text to speech. arXiv preprint
arXiv:2006.04558.
Schweter, S. (2020). Berturk - bert models for turkish.
Stepanov, I. and Shtopko, M. (2024). Gliner multi-task:
Generalist lightweight model for various information
extraction tasks. arXiv preprint arXiv:2406.12925.
Yamamoto, R., Song, E., and Kim, J.-M. (2020). Parallel
wavegan: A fast waveform generation model based on
generative adversarial networks with multi-resolution
spectrogram. In ICASSP 2020-2020 IEEE Interna-
tional Conference on Acoustics, Speech and Signal
Processing (ICASSP), pages 6199–6203. IEEE.
Zen, H., Dang, V., Clark, R., Zhang, Y., Weiss, R. J., Jia,
Y., Chen, Z., and Wu, Y. (2019). Libritts: A cor-
pus derived from librispeech for text-to-speech. arXiv
preprint arXiv:1904.02882.
An End-to-End Generative System for Smart Travel Assistant
479