1 of 30

Deepfake Word Detection by Next-token Prediction using Fine-tuned Whisper

Hoan My Tran1∗, Xin Wang2∗, Wanying Ge2, Xuechen Liu2, Junichi Yamagishi2

1Université de Rennes, France

2National Institute of Informatics, Japan

Interspeech 2026

Area 4 - Oral 2 - 3

wangxin@nii.ac.jp

1

2 of 30

Slides by Xin Wang

National Institute of Informatics

​

​

​

© 2026, Xin Wang. All rights reserved.

​

​

​

​

This work is licensed under the Creative Commons Attribution 3.0 license.

See http://creativecommons.org/ for details.

​

​

​

​

Note: Speech data is from ASVspoof 2019 database https://datashare.ed.ac.uk/handle/10283/3336

2

3 of 30

One-minute summary

  • Finetune a Whisper to detect fake words via next-token prediction
    • ASR + fake words detection at the same time
    • Method
      1. Prepare the finetuning data
        • Vocoding randomly selected words
        • Add <special token> around vocoded words in target token sequences
      2. Finetune Whisper

​

  • Good results on in-domain test sets
  • Improvement is needed on out-of-domain data

​

3

4 of 30

Introduction

4

5 of 30

Introduction

Example from Meta VoiceCraft: https://jasonppy.github.io/VoiceCraft_web/

https://www.bloomberg.com/news/articles/2024-01-26/ai-startup-elevenlabs-bans-account-blamed-for-biden-audio-deepfake?embedded-checkout=true

  • Popular task: whether an utterance is fake

Detector

Deepfake generator

REAL

FAKE

Attacker

Defender

Transmission

VoIP

mp3

or …

They knew that government don’t money directly

5

6 of 30

Introduction

Example from Meta VoiceCraft: https://jasonppy.github.io/VoiceCraft_web/

https://www.bloomberg.com/news/articles/2024-01-26/ai-startup-elevenlabs-bans-account-blamed-for-biden-audio-deepfake?embedded-checkout=true

  • Task in this paper: which part is fake?

?

Deepfake generator

Transmission

VoIP

mp3

or …

real voice

M

mix real & fake voices

They knew that government don’t money directly

Attacker

Defender

6

7 of 30

Introduction

Example from Meta VoiceCraft: https://jasonppy.github.io/VoiceCraft_web/

https://www.bloomberg.com/news/articles/2024-01-26/ai-startup-elevenlabs-bans-account-blamed-for-biden-audio-deepfake?embedded-checkout=true

  • Task in this paper: which part is fake?

Fake word detector

Deepfake generator

Transmission

VoIP

mp3

or …

real voice

M

mix real & fake voices

Attacker

Defender

They knew that government don’t money directly

They knew that government don’t money directly

7

8 of 30

Introduction

  • Existing method

Fake word detector

They knew that government don’t money directly

Fake interval detector

Real

Fake

Speech recognizer

They knew that government don’t money directly

Feature extraction

(SSL, filter-bank.)

Classifier

(ResNet, LCNN)

Real

Fake

Simpler solution to transcribe the speech & locate fake words?

Spoofing localization

8

9 of 30

Introduction

  • Our method

Fake word detector

They knew that government don’t money directly

Whisper

They knew that government don’t <TOF>money directly<EOF>

Next-token prediction

This is an

<TOF>example<EOF>

Pseudo fake (vocoded) dataset

Fine-tuning

9

10 of 30

Method

10

11 of 30

Recap: Whisper

Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. In Proc. ICML, 2023. PMLR, 28492–28518.

​

This

is

a

Training dataset

This is a good example!

Waveform

Text tokens

11

12 of 30

Proposed method – at the high level

Training dataset

This is a good example!

This

is

a

Waveform partially vocoded

Text tokens

Prepare vocoded training data

Fine-tune Whisper

(using default recipe)

This is a<TOF>good<EOF> example!

Pseudo fake (vocoded)

12

13 of 30

Proposed method – prepare data

See code pipeline: https://github.com/nii-yamagishilab/project-NN-Pytorch-scripts/tree/master/package

This is a good example!

Training dataset

This is a good example!

Waveform

Text tokens

(text transcription)

  1. Force-align text & waveform
  2. Randomly pick one to three words
  3. Vocode the word using a vocoder & overlap-add
  4. Add special tokens

13

14 of 30

Proposed method – prepare data

See code pipeline: https://github.com/nii-yamagishilab/project-NN-Pytorch-scripts/tree/master/package

This is a good example!

Training dataset

This is a good example!

Waveform

Text tokens

(text transcription)

  1. Force-align text & waveform
  2. Randomly pick one to three words
  3. Vocode the word using a vocoder & overlap-add
  4. Add special tokens

14

15 of 30

Proposed method – prepare data

See code pipeline: https://github.com/nii-yamagishilab/project-NN-Pytorch-scripts/tree/master/package

This is a good example!

Training dataset

This is a good example!

Waveform

Text tokens

(text transcription)

  1. Force-align text & waveform
  2. Randomly pick one to three words
  3. Vocode the word using a vocoder & overlap-add
  4. Add special tokens

Vocoding: analysis & synthesis

15

16 of 30

Proposed method – prepare data

See code pipeline: https://github.com/nii-yamagishilab/project-NN-Pytorch-scripts/tree/master/package

Training dataset

This is a good example!

Waveform

Text tokens

This is a good example!

  1. Force-align text & waveform
  2. Randomly pick one to three words
  3. Vocode the word using a vocoder & overlap-add
  4. Add special tokens

(text transcription)

Vocoding: analysis & synthesis

16

17 of 30

Proposed method – prepare data

See code pipeline: https://github.com/nii-yamagishilab/project-NN-Pytorch-scripts/tree/master/package

Training dataset

This is a good example!

Waveform

Text tokens

(text transcription)

This is a<TOF>good<EOF> example!

<TOF> !!!!!!

<EOF> ~~~

We re-use existing tokens:

  1. Force-align text & waveform
  2. Randomly pick one to three words
  3. Vocode the word using a vocoder & overlap-add
  4. Add special tokens

Vocoding: analysis & synthesis

17

18 of 30

Proposed method – fine-tuning Whisper

We tried LoRA, or fine-tune encoder / decoder only, but performance was sub-optimal

This

is

a

Our best setup:

just fine-tuning the whole Whisper

Training dataset

This is a<TOF>good<EOF> example!

18

19 of 30

Examples of Whisper’s prediction

!!!!!!Now~~~ this !!!!!!isn't~~~ absolutely definitive.

!!!!!!Now~~~ this !!!!!!isn't~~~ absolutely definitive.

​

​

It seemed a moving and !!!!!!fitting~~~ !!!!!!destruction.~~~

It seemed a moving and fitting !!!!!!distraction.~~~

​

​

​

I really !!!!!!didn’t~~~ enjoy putting ...

I really !!!!!!didn’t~~~ !!!!!!kind~~~ of putting ...

Ground-truth:

Whisper:

False acceptance (of a fake word)

False rejection (of a real word)

Ground-truth:

Whisper:

Ground-truth:

Whisper:

Dataset: E.VOC

Dataset: PartialEdit

Dataset: AVDeepfake-1m

19

20 of 30

Experiment

20

21 of 30

Experiment setup

  • Research questions
    • Can fine-tuned Whisper detect synthetic words while transcribing?
    • Is the vocoding-based training data helpful?

​

  • Evaluation metrics
    • False rejection rate (FRR, real words misclassified as fake)
    • False acceptance rate (FAR, fake words misclassified as real)
    • Word error rates (WER)

​

​

​

​

​

​

​

​

​

      • e.g., given word alignment, FRR =

​

#. real words in eva. set

#. misclassified to be fake

21

22 of 30

Experiment setup

Hieu-Thi Luong, Haoyang Li, Lin Zhang, Kong Aik Lee, and Eng Siong Chng. 2025. LlamaPartialSpoof: An LLM-Driven Fake Speech Dataset Simulating Disinformation Generation. In Proc. ICASSP.

  • Research questions
    • Can fine-tuned Whisper detect synthetic words while transcribing?
    • Is the vocoding-based training data helpful?

​

  • Data

​

Vocoded data

Text-to-speech (TTS)-based data

Llamapartialspoof (Luong 2025)

Mix of both

Vocoder

Domain: LibriVox

Text-to-speech model

22

23 of 30

Experiment setup

Hieu-Thi Luong, Haoyang Li, Lin Zhang, Kong Aik Lee, and Eng Siong Chng. 2025. LlamaPartialSpoof: An LLM-Driven Fake Speech Dataset Simulating Disinformation Generation. In Proc. ICASSP.

  • Research questions
    • Can fine-tuned Whisper detect synthetic words while transcribing?
    • Is the vocoding-based training data helpful?

​

  • Data

​

Vocoded test data

TTS test data from Llamapartialspoof (Luong 2025)

Out-of-domain TTS data

23

24 of 30

Results – Audiobook test data

Nicholas Klein, Hemlata Tak, James Fullwood, Krishna Regmi, Leonidas Spinoulas, Ganesh Sivaraman, Tianxiang Chen, and Elie Khoury. 2025. Pindrop it! Audio and visual deepfake countermeasures for robust detection and fine-grained localization. In Proc. ACM MM (Mm ’25), 2025. Association for Computing Machinery, Dublin, Ireland, 13700–13706.

​

​

WER (%)

FAR (%)

FRR (%)

ResNet-152

(AVDeepfake 1M++ top-1, Klein 2025)

Evaluate on vocoded data (E.VOC)

Fine-tuned / trained using

vocoded data Ft.VOC

0.87

-

-

-

7.22

7.15

0.52

3.81

Pre-trained Whisper

23.89

The lower the better

large v3

Whisper

large v3

RQ1: it is possible that a fine-tuned Whisper can detect synthetic (vocoded) words while transcribing

24

25 of 30

Results – Audiobook test data

FAR (%)

FRR (%)

ResNet-152

(AVDeepfake 1M++ top-1, Klein 2025)

Evaluate on vocoded data (E.VOC)

Fine-tuned / trained using

vocoded data Ft.VOC

7.22

7.15

0.52

3.81

Whisper

large v3

Evaluate on TTS data (E.TTS)

FAR (%)

FRR (%)

9.01

10.14

76.16

83.04

The lower the better

☹

Detecting unseen synthetic methods?

Most of the fake words are mis-classified as REAL

25

26 of 30

Results – Audiobook test data

FAR (%)

FRR (%)

Whisper

ResNet-152

(AVDeepfake 1M++ top-1, Klein 2025)

Evaluate on vocoded data (E.VOC)

Fine-tuned / trained using

vocoded data Ft.VOC

7.22

7.15

0.52

3.81

large v3

RQ2: cross-dataset generalization is an issue for both models

Evaluate on TTS data (E.TTS)

FAR (%)

FRR (%)

9.01

10.14

76.16

83.04

The lower the better

☹

Whisper

ResNet-152

Fine-tuned / trained using

TTS data Ft.TTS

large v3

18.86

76.42

78.59

9.08

1.38

0.15

1.79

3.13

☹

26

27 of 30

Summary

27

28 of 30

Summary

  • Detecting fake words via next-token prediction
    • create vocoded training data & add special tokens
    • fine-tune Whisper

Fine-tuned Whisper

They knew that government don’t money directly

Vocoded data

This is a<TOF>good<EOF> example!

fine-tuning

  • It is feasible to detect synthetic words via next-token prediction
  • Room for improvement
    • Unseen attacks
    • Unseen domains (Youtube, studio) 🡪 see paper
    • Diverse training data may be the key

28

29 of 30

Related work

Jiangyan Yi, Ruibo Fu, Jianhua Tao, Shuai Nie, Haoxin Ma, Chenglong Wang, Tao Wang, Zhengkun Tian, Ye Bai, Cunhang Fan, Shan Liang, Shiming Wang, Shuai Zhang, Xinrui Yan, Le Xu, Zhengqi Wen, and Haizhou Li. 2022. ADD 2022: the first Audio Deep Synthesis Detection Challenge. In Proc. ICASSP, May 2022. 9216–9220.

Yupei Li, Li Wang, Yuxiang Wang, Lei Wang, Rizhao Cai, Jie Shi, Björn W. Schuller, and Zhizheng Wu. 2025. DFALLM: Achieving Generalizable Multitask Deepfake Detection by Optimizing Audio LLM Components.

Lin Zhang, Xin Wang, Erica Cooper, Junichi Yamagishi, Jose Patino, and Nicholas Evans. 2021. An Initial Investigation for Detecting Partially Spoofed Audio. In Proc. Interspeech, August 2021. 4264–4268.

Lin Zhang, Xin Wang, Erica Cooper, Nicholas Evans, and Junichi Yamagishi. 2022. The PartialSpoof Database and Countermeasures for the Detection of Short Fake Speech Segments Embedded in an Utterance. IEEE/ACM Transactions on Audio, Speech, and Language Processing (2022), 1–13.

Lin Zhang, Xin Wang, Erica Cooper, Nicholas Evans, and Junichi Yamagishi. 2023. Range-Based Equal Error Rate for Spoof Localization. In Proc. Interspeech, August 20, 2023. ISCA, 3212–3216.

Piotr Kawa, Marcin Plata, Michał Czuba, Piotr Szymański, and Piotr Syga. 2023. Improved DeepFake Detection Using Whisper Features. In Proc. Interspeech, August 20, 2023. ISCA, 4009–4013.

Lin Zhang, Thomas Thebaud, Zexin Cai, Sanjeev Khudanpur, Daniel Povey, Leibny Paola García-Perera, Matthew Wiesner, and Nicholas Andrews. 2026. Can LLMs Help Localize Fake Words in Partially Fake Speech?

Hieu-Thi Luong, Haoyang Li, Lin Zhang, Kong Aik Lee, and Eng Siong Chng. 2025. LlamaPartialSpoof: An LLM-Driven Fake Speech Dataset Simulating Disinformation Generation. In Proc. ICASSP, January 06, 2025. arXiv.

Nicholas Klein, Hemlata Tak, James Fullwood, Krishna Regmi, Leonidas Spinoulas, Ganesh Sivaraman, Tianxiang Chen, and Elie Khoury. 2025. Pindrop it! Audio and visual deepfake countermeasures for robust detection and fine-grained localization. In Proc. ACM MM (Mm ’25), 2025. Association for Computing Machinery, Dublin, Ireland, 13700–13706.

  • Partial fake (interval) detection
    • Metrics: counting (Yi 2022), intersection of union (Li 2025), range-EER (Zhang 2023)
    • Models: single-resolution (Yi 2022, Zhang 2022, Klein 2025), multi-resolution (Zhang 2022)
    • Datasets: PartialSpoof (Zhang 2022), LlamaPartialSpoof (Luong 2025), PartialEdit (Zhang 2025)

​

  • Using Whisper or large language model for fake detection
    • Whisper as feature extractor for utterance-level detection (Kawa 2023)
    • Large language model (LLM) fake word detection (Lin 2026)
    • Many other latest work using LLMs

​

29

30 of 30

Thank you

Code & checkpoint

https://github.com/nii-yamagishilab/Whisper-deepfake-word-detection

30