Open Vocabulary Keyword Spotting through Transfer Learning from Speech Synthesis

Kesavaraj V, Anil Kumar Vuppala
Speech Processing Laboratory, LTRC,
International Institute of Information Technology Hyderabad, India
kesavaraj.v@research.iiit.ac.in, anil.vuppala@iiit.ac.in

Abstract

Identifying keywords in an open-vocabulary context is crucial for personalizing interactions with smart devices. Previous approaches to open vocabulary keyword spotting depend on a shared embedding space created by audio and text encoders. However, these approaches suffer from heterogeneous modality representations (i.e., audio-text mismatch). To address this issue, our proposed framework leverages knowledge acquired from a pre-trained text-to-speech (TTS) system. This knowledge transfer allows for the incorporation of awareness of audio projections into the text representations derived from the text encoder. The performance of the proposed approach is compared with various baseline methods across four different datasets. The robustness of our proposed model is evaluated by assessing its performance across different word lengths and in an Out-of-Vocabulary (OOV) scenario. Additionally, the effectiveness of transfer learning from the TTS system is investigated by analyzing its different intermediate representations. The experimental results indicate that, in the challenging LibriPhrase Hard dataset, the proposed approach outperformed the cross-modality correspondence detector (CMCD) method by a significant improvement of 8.22% in area under the curve (AUC) and 12.56% in equal error rate (EER).

Index Terms—Transfer learning, Text-to-Speech, Keyword spotting, Tacotron 2

I Introduction

Keyword spotting (KWS) is a process of detecting specific keywords within a continuous audio stream, which is crucial for enabling voice-driven interactions on edge devices [1, 2]. Increasing demand for personalized voice assistants highlights the significance of user-defined keyword spotting, which is known as custom keyword detection or open vocabulary keyword spotting [3, 4]. In contrast to closed vocabulary keyword spotting [5], where only predetermined keywords are recognized, custom keyword spotting, deals with the difficulty of identifying random keywords that the model might not have been exposed to during training, introducing an additional level of complexity to the task.

In the literature, there exist different methods for custom keyword spotting. One such method is the query-by-example (QbyE) [6, 7] approach, which involves matching input queries with pre-enrolled examples. However, the effectiveness of the QbyE method relies heavily on the similarity between the recorded speech during enrollment and the subsequent evaluated speech recordings. Factors like diverse vocal characteristics among users and environments with background noise pose significant challenges that can impact the consistency of performance of the QbyE method. In addressing this concern, researchers have delved into text enrollment-based methods [3, 8]. One such method is the automatic speech recognition-based approach [9], which detects phonetic patterns in input streams and then compares them with enrolled keyword representation. However, the effectiveness of this method heavily depends on the accuracy of the acoustic model. In [8], an attention-based cross-modal matching approach is proposed that is trained in an end-to-end manner with monotonic matching loss and keyword classification loss. In [10], a zero-shot KWS is proposed, coupled with phoneme-level detection loss. Also, [11] introduced Dynamic Sequence Partitioning to optimally partition the audio sequence to match the length of the word-based text sequence. These recent end-to-end techniques [10, 11, 8], primarily depend on evaluating speech and text representations in a common latent space, demonstrated promising results in the custom keyword spotting task. Therefore, we considered the end-to-end approach for our investigation.

Refer to caption — Figure 1: Proposed architecture for open vocabulary keyword spotting

The effectiveness of the end-to-end approaches relies on the accuracy of projecting audio and text representations into a shared embedding space. While these methods [8, 10] project speech and text representations into a shared latent space, they encounter difficulties in distinguishing between pairs of closely related pronunciations. This challenge can be addressed by generating text representations that have some acoustic knowledge. On the other hand, a TTS model converts text to audio, and the intermediate representations derived from this process possess an understanding of audio projections. Consequently, insights acquired from a pre-trained TTS model can serve as meaningful representations of text. In [12], a similar approach is used to transfer knowledge from a pre-trained TTS model for voice conversion task. Hence, in this paper, a novel framework is introduced to distill knowledge from a pre-trained TTS model for open-vocabulary keyword spotting task. The contributions of this study can be summarized as follows:

•

A novel strategy is proposed that leverages intermediate representations extracted from a pre-trained TTS model as valuable text representations for custom KWS tasks.
•

Extensive investigation is conducted on various outputs of intermediate layers of the pre-trained TTS model to assess the efficacy of transfer learning.
•

An ablation study is conducted to examine the system’s performance for keywords of different word lengths.
•

Additionally, the robustness of the proposed framework is analyzed in an OOV scenario.

The following sections of the paper are organized as follows: Section II introduces the proposed method in this study. Section III provides details about the experimental setup, Section IV presents the results and discussion, and Section V concludes the study.

II Proposed Method

This study introduces a novel methodology for open vocabulary keyword spotting by leveraging knowledge from a pre-trained TTS model. In this section, we describe the proposed architecture, as shown in Fig. 1. The architecture consists of four submodules: text encoder, audio encoder, pattern extractor, and pattern discriminator. The following sections contain detailed information about individual modules.

II-A Text Encoder

It consists of a pre-trained Tacotron 2 [13] model, a recurrent sequence-to-sequence TTS system, which takes character sequence as input for the corresponding keyword. The resulting intermediate representations from the TTS model are then directed to a bidirectional gated recurrent unit (Bi-GRU) layer with a dimension of 64. The dimensions of the intermediate representations from the TTS model depend on the specific layer from which they are extracted. Further details of the intermediate representations are discussed in subsection III-B. The output from the Bi-GRU layer is then fed to a dense layer of size 128 units. The output from the text encoder is denoted as $T_{E}\in\mathbb{R}^{d\times m}$ , where d and m denote the embedding dimension and length of the text (i.e. number of characters), respectively. The primary motivation for incorporating the pre-trained TTS system into the text encoder is to generate text representations that are aware of acoustic projections. This integration simplifies the task of projecting audio and text embeddings in shared latent space. Although this strategy does not explicitly involve generating audio from text, but exploits knowledge transfer from intermediate representations of TTS.

II-B Audio Encoder

The input to the audio encoder is 80-dimensional mel-filterbank coefficients, extracted for every 10ms with a 25ms window length. To process this audio feature, the encoder consists of two 2-D convolution layers (Conv 2D) with a kernel size of 3. The first convolution layer consists of 32 filters, and the second convolution layer consists of 64 filters. Batch normalization is applied after each convolution layer to ensure stable training. To improve computational efficiency, the initial convolutional layer employs a stride of 2 to effectively reduce the number of processed frames by skipping consecutive frames. Following the convolution layers, two Bi-GRU layers each having a dimension of 64 are employed. Subsequently, the output from the last Bi-GRU layer is passed into a dense layer which generates a 128-dimensional audio embedding. The output from the audio encoder is denoted as $A_{E}\in\mathbb{R}^{d\times n}$ , where d and n denote the embedding dimension and length of the audio (i.e. number of frames), respectively.

TABLE I: Performance comparison of the proposed method with various KWS techniques across different datasets: Google Commands V1 (G), Qualcomm Keyword Speech dataset (Q), LibriPhrase-Easy (LP_E), and LibriPhrase-Hard (LP_H).

Method	EER (%)				AUC (%)
Method	G	Q	LP_E	LP_H	G	Q	LP_E	LP_H
CTC [7]	31.65	18.23	14.67	35.22	66.36	89.69	92.29	69.58
Attention [6]	14.75	49.13	28.74	41.95	92.09	50.13	78.74	62.65
Triplet [3]	35.6	38.72	32.75	44.36	71.48	66.44	63.53	54.88
CMCD [8]	27.25	12.15	8.42	32.9	81.06	94.51	96.7	73.58
Proposed	22.3	10.82	5.61	24.68	85.16	95.65	98.49	86.14

II-C Pattern Extractor

It is based on the cross-attention mechanism [14] which captures the temporal correlations between the audio and text embeddings. In this setup, the audio embedding $A_{E}$ functions as both the key and value, while the text embedding $T_{E}$ acts as the query. The output of the pattern extractor is the context vector which contains information about audio and text agreement.

II-D Pattern Discriminator

It consists of a single Bi-GRU layer with a dimension of 128 that takes the context vector from the cross-attention layer as input. Output from the last frame of the Bi-GRU layer is fed into a dense layer with a sigmoid as an activation function. The pattern discriminator determines whether audio and text inputs share the same keyword or not.

III Experimental Setup

This section describes the database, Tacotron 2 embeddings, and implementation details for training.

III-A Database

For training and evaluation, we used the Libriphrase [8] dataset, which comprises short phrases with varying word lengths (1 to 4). The dataset was generated from the Librispeech corpus [15]. The training set of Libriphrase is generated using train-clean-100 and train-clean-360 subsets, and the evaluation set is using train-others-500 subset. The evaluation set consists of 4391, 2605, 467, and 56 episodes of each word length respectively. Each episode has three positive and three negative pairs. The negative samples are further categorized into easy and hard based on Levenshtein distance [16], leading to the creation of the Libriphrase Easy (LP_E) and Libriphrase Hard (LP_H) datasets. Each sample in the dataset is represented by using three entities: audio, text, and a binary target value indicating 1 for a positive pair and 0 for a negative pair.

To comprehensively evaluate model performance, we extended our assessment beyond the Libriphrase dataset by incorporating the Google Speech Commands V1 dataset (G) [17] and the Qualcomm Keyword Speech dataset (Q) [18]. The Google Speech Commands V1 dataset (G) encompasses speech recordings from 1,881 speakers, focusing on 30 small keywords. On the other hand, the Qualcomm Keyword Speech dataset (Q) comprises 4,270 utterances of four keywords spoken by 50 speakers.

III-B Tacotron 2 Embeddings

In this study, the utilization of intermediate representation from Tacotron 2 is proposed to enhance performance, particularly in scenarios involving speech-keyword pairs with similar pronunciations. The intermediate representations are obtained using NVIDIA Tacotron 2 model [19] which is pre-trained on the LJSpeech dataset [20]. The details of different intermediate representations that are used in this study are outlined in Table II. Additional insights into the Tacotron 2 architecture can be found in [13]

TABLE II: Description of intermediate representations of Tacotron 2 model.

T_{a}-\text{number of characters}

T_{b}-\text{number of frames derived from the Tacotron 2}

Embedding	Dimension	Description
E1	$T_{a}\times 512$	CharEmbedding block output
E2	$T_{a}\times 512$	Convolution block output
E3	$T_{a}\times 512$	Bi-LSTM block output
E4	$T_{b}\times 512$	Attention block output
E5	$T_{b}\times 512$	Prenet block output
E6	$T_{b}\times 512$	Postnet block output
E7	$T_{b}\times 80$	Target Melspectrogram

TABLE III: Evaluation of different intermediate representations of Tacotron 2 model for custom KWS task across different datasets: Google Commands V1 dataset (G), Qualcomm Keyword Speech dataset (Q), LibriPhrase-Easy dataset (LP_E), and LibriPhrase-Hard dataset (LP_H)

Method	EER (%)				AUC (%)				F1 score (%)
Method	G	Q	LP_E	LP_H	G	Q	LP_E	LP_H	G	Q	LP_E	LP_H
E1	31.44	19.32	9.26	32.13	74.84	88.77	96.29	74.67	69.23	81.45	91.2	64.86
E2	26.43	14.52	7.11	28.47	80.66	93.02	97.67	79.6	74.2	79.83	93.26	67.78
E3	22.3	10.82	5.61	24.68	85.16	95.65	98.49	86.14	78.56	87.74	94.5	72.2
E4	24.6	15.27	8.43	31.2	81.83	90.63	96.61	76.25	74.77	68.23	91.7	67.54
E5	27.51	19.98	12.02	34.35	79.71	87.55	94.5	71.77	66.42	72.58	86.94	71.41
E6	34.61	33.14	20.9	39.19	69.95	72.81	87.06	65.29	66.32	66.85	78.51	61.47
E7	25.42	23.79	11.68	34.7	81.14	84.5	94.57	70.97	73.14	71.2	87.5	66.78

III-C Implementation Details

The training pipeline is structured as a binary classification task, wherein the model is tasked with classifying the similarity of input pairs {text, audio}. In audio and text encoders, weights are initialized through Xavier initialization, and Leaky ReLU is utilized as the activation function. A dropout of 0.2 is added after each layer in the encoders to prevent overfitting. The training process uses binary cross-entropy loss as the training criterion and employs the Adam optimizer [21] with default parameters for optimization. A fixed learning rate of $10^{-4}$ and a batch size of 128 is employed in the training process. The best-performing model was chosen based on its performance on the validation set. For training, we used four NVIDIA GeForce RTX 2080 Ti GPUs.

IV Results and Discussion

The proposed method explores intermediate representations from a pre-trained TTS model to boost the initialization of the text encoder. By tapping into the audio projections within the pre-trained TTS model, we aim to enrich text representations for custom keyword spotting task. We conducted extensive experiments across diverse datasets to evaluate this approach, with results presented in Tables - I, III, IV, V.

Table I presents the comparative analysis of baselines and proposed approach across G, Q, LP_E, and LP_H datasets. Evaluation results show that among all the baselines CMCD demonstrates strong performance and Triplet shows weak performance across the Q, LP_E, and LP_H datasets, in terms of AUC and EER. On the other hand, the attention-based QbyE method shows powerful performance on the G dataset, which consists of frequently used words (e.g., ”on”, ”off”) as keywords. This method benefits when the keyword is part of the training set due to its similarity scoring mechanism, but it shows degraded performance when the keyword is unfamiliar, as observed in Q and LibriPhrase. However, our proposed approach outperforms all the baselines on all datasets except G. It is also evident, in comparison with the CMCD baseline, the proposed method showcases a significant improvement of 8.22 % in AUC and 12.56% in EER on the challenging LP_H dataset which consists of similar audio-text pairs (e.g., ”madame” and ”modem”). This substantial improvement highlights the effectiveness of our method in better discrimination of closely related pronunciations. Moreover, we measure the generalization of the model on G and Q datasets, without any finetuning. We find a consistent improvement of around 3% on the AUC metric and 2.62% on the EER metric across G and Q datasets when compared with the CMCD baseline.

Further, to assess the efficacy of transfer learning from pre-trained TTS for custom KWS task, an ablation study is conducted. This study compares embeddings obtained from various intermediate layers of the Tacotron-2 model, and results of this comparison are reported in Table III. Analyzing the results, E3 consistently outperforms others in terms of lower Equal Error Rate (EER) and higher AUC and F1-score across all datasets. This suggests it captures both acoustic and linguistic information of the enrolled keyword more effectively. Moreover, E4 and E2 exhibit competitive performance with consistently good AUC scores, signifying them as better alternatives to E3. Conversely, E6 shows poor performance compared to other layer embeddings, indicating its limitations for this task. Additionally, it can be inferred that intermediate representations (E2, E3) from the Tacotron 2 encoder seem to be significantly more suitable for the custom KWS task in comparison to representations (E5, E6, E7) from the Tacontron 2 decoder. This implies that capturing information before mel-spectrogram generation in the encoder stage is crucial for accurate keyword detection.

TABLE IV: Performance of the proposed approach across different word lengths

Word length	EER (%)	AUC (%)	F1 score (%)
1	5.41	98.07	94.46
2	5.9	97.83	94.24
3	7.59	97.04	92.28
4	8.5	97.12	90.85

Table IV presents the performance analysis of the proposed system across various word lengths. The evaluation of a system across different word lengths (1 to 4) reveals that shorter words (1 or 2) result in better performance, with lower EER and higher F1-score suggesting higher accuracy in keyword identification. As word length increases, EER values rise, indicating increased difficulty in recognition. However, our method exhibits consistent performance across all word lengths. Additionally, to assess the robustness of the proposed system, we compared its performance in the OOV scenario. From Table V, it is evident that, in comparison to the CMCD baseline method, the proposed method shows an absolute improvement of 7.25%, 6.36%, and 5.53% in terms of F1 score, AUC and EER, respectively. This signifies the effectiveness of the proposed system in handling user-defined keywords that are not seen during training.

TABLE V: Comparison of CMCD and proposed approach in Out-of-Vocabulary scenario.

Method	EER (%)	AUC (%)	F1 score (%)
CMCD	23.48	84.08	76.2
Proposed	18.14	90.44	83.45

Additionally, an interesting observation was made that the proposed network, utilizing intermediate representations extracted from E3, exhibits faster convergence, as illustrated in Fig. 2.

V Conclusion

This study presented an end-to-end architecture for open vocabulary keyword spotting that leverages insights from a pre-trained TTS system. The proposed approach exhibited competitive performance compared to baseline methods across four different datasets. Notably, it showcases its potential in distinguishing similar pronunciations of audio-text pairs in the Libriphrase hard dataset. The ablation study on intermediate layers of the Tacotron 2 model revealed that E3 (Bi-LSTM block output) exhibited the best performance and faster convergence during training. Moreover, the proposed approach showed consistent performance in keyword identification regardless of the word lengths considered and demonstrated its robustness in the OOV condition. In future work, the focus will be on optimizing the knowledge transfer effectively by exploring effective strategies to utilize all intermediate layers of the TTS model rather than relying on a single layer.

References

[1] Iván López-Espejo, Zheng-Hua Tan, John HL Hansen and Jesper Jensen “Deep spoken keyword spotting: An overview” In IEEE Access 10 IEEE, 2021, pp. 4169–4199
[2] Raphael Tang and Jimmy Lin “Deep residual learning for small-footprint keyword spotting” In Proc. ICASSP, 2018, pp. 5484–5488 IEEE
[3] Niccolo Sacchi, Alexandre Nanchen, Martin Jaggi and Milos Cernak “Open-vocabulary keyword spotting with audio and text embeddings” In INTERSPEECH 2019-IEEE International Conference on Acoustics, Speech, and Signal Processing, 2019
[4] Krishna Gurugubelli, Sahil Mohamed and Rajesh Krishna KS “Comparative Study of Tokenization Algorithms for End-to-End Open Vocabulary Keyword Detection” In Proc. ICASSP, 2024, pp. 12431–12435 IEEE
[5] Tara N. Sainath and Carolina Parada “Convolutional neural networks for small-footprint keyword spotting” In Proc. Interspeech, 2015, pp. 1478–1482
[6] Jinmiao Huang, Waseem Gharbieh, Han Suk Shim and Eugene Kim “Query-by-example keyword spotting system using multi-head attention and soft-triple loss” In Proc. ICASSP, 2021, pp. 6858–6862 IEEE
[7] L. Lugosch, S. Myer and V.. Tomar “DONUT: CTC-based Query-by-Example Keyword Spotting” In NeurIPS Workshop on Interpretability and Robustness in Audio, Speech, and Language, 2018
[8] Hyeon-Kyeong Shin et al. “Learning Audio-Text Agreement for Open-vocabulary Keyword Spotting” In Proc. INTERSPEECH, 2022, pp. 1871–1875
[9] Zuozhen Liu, Ta Li and Pengyuan Zhang “Neural keyword confidence estimation for open-vocabulary keyword spotting” In Electronics Letters 58.3 IET, 2021, pp. 133–135
[10] Yong-Hyeok Lee and Namhyun Cho “PhonMatchNet: Phoneme-Guided Zero-Shot Keyword Spotting for User-Defined Keywords” In Proc. INTERSPEECH, 2023, pp. 3964–3968
[11] Kumari Nishu, Minsik Cho and Devang Naik “Matching Latent Encoding for Audio-Text based Keyword Spotting” In Proc. INTERSPEECH, 2023, pp. 1613–1617
[12] Wen-Chin Huang et al. “Voice Transformer Network: Sequence-to-Sequence Voice Conversion Using Transformer with Text-to-Speech Pretraining” In Proc. Interspeech, 2020, pp. 4676–4680
[13] Jonathan Shen et al. “Natural tts synthesis by conditioning wavenet on mel spectrogram predictions” In Proc. ICASSP, 2018, pp. 4779–4783 IEEE
[14] Ashish Vaswani et al. “Attention is all you need” In Advances in neural information processing systems 30, 2017
[15] Vassil Panayotov, Guoguo Chen, Daniel Povey and Sanjeev Khudanpur “Librispeech: an asr corpus based on public domain audio books” In Proc. ICASSP, 2015, pp. 5206–5210 IEEE
[16] Vladimir I Levenshtein “Binary codes capable of correcting deletions, insertions, and reversals” In Soviet physics doklady 10.8, 1966, pp. 707–710 Soviet Union
[17] Pete Warden “Speech commands: A dataset for limited-vocabulary speech recognition” In arXiv preprint arXiv:1804.03209, 2018
[18] Byeonggeun Kim et al. “Query-by-example on-device keyword spotting” In Proc. ASRU, 2019, pp. 532–538 IEEE
[19] NVIDIA Corporation “Pretrained Tacotron2 model”, https://github.com/NVIDIA/tacotron2
[20] Keith Ito “The LJ Speech Dataset”, https://keithito.com/LJ-Speech-Dataset/, 2017
[21] Diederik P Kingma and Jimmy Ba “Adam: A method for stochastic optimization” In arXiv preprint arXiv:1412.6980, 2014