Please note: I am writing this research paper summaries in mind to build an advanced(State of the art) Text-to speech system for Malayalam, as a FOSS project for SMC. I would like to thank for all the help I have recieved from Santhosh Thottingal, so far and suggesting me this project.

Abstract:

Currently, there is an increasing interests in text-to-speech (TTS) synthesis to use sequence-to-sequence models with attention. However, in challenging speaking styles, like Lombard speech, which has higher intensity and fundamental frequency(F0) being large, cuurent approaches are not always efficent. In this study we propose a transfer learning method to adapt a sequence-to-sequence based TTS system of normal speaking style to Lombard style. More- over, we experiment with a WaveNet vocoder in synthesis of Lombard speech. We conducted subjective evaluations to assess the per- formance of the adapted TTS systems. The subjective evaluation results indicated that an adaptation system with the WaveNet vocoder clearly outperformed the conventional deep neural network based TTS system in synthesis of Lombard speech.

A system which was build using the Seq2Seq-TTS models and the mel-spectogram as output which employed the Wavenet Vocoder got higher accuracy when trained with only 30 min- utes of Lombard speech, system S5 generated synthetic speech that was most Lombard-like among the systems compared.

The contributions of the paper are twofold. First, we develop a speaking style adaptation system using a Seq2Seq-TTS model. Sec- ond, we study the use of a WaveNet vocoder for the application of Lombard speech synthesis. To the best of our knowledge, the cur- rent study is the first investigation on speaking style adaptation of speech synthesis using a modern Seq2Seq-TTS system.

SEQ2SEQ-TTS System

Despite promising results in synthetic speech using Statistic parametric speech synthesis. These models depend heavily on encoder-decoder neural network structures that map a sequence of characters to a sequence of acous- tic frames. These models combine the front-end and back-end and learn relations between them from data only. When sequence-to- sequence models are coupled with neural vocoders, they enable generating raw waveforms directly from text. It was demonstrated that state-of-the-art results in TTS can be achieved with the sequence-to-sequence technology.

The model accepts either mono-phonemes or graphemes as inputs and emits acoustic parameters as outputs. It consists of three main components: 1) en- coder, 2) attention, and 3) decoder. The encoder takes text sequence x of length L as input, which represented either in the character or phoneme domain as one-hot vectors. The encoder learns a continu- ous sequential representation h using various neural network archi- tectures such as LSTMs and/or CNNs.

h = encoder(x)

αt = attention(st−1 , αt−1 , h)

ct = sum of(αt, ht)

yt = decoder(st−1 , ct )
Equations in paper to fear you

where st−1 is the (t − 1)-th state of the decoder recurrent neural network and αt ∈ RL are the attention weights or the alignment and ct is the context or attention vector. The decoder takes the previous hidden state st−1 and the current context vector ct as inputs and generates the current output yt . This process runs until the end of the utterance is reached

Adaption of Seq2Seq models

Using Adaption we are able to generate amamzing results that show models which generate synthetic speech of good quality using around 30 minutes of data. In the present study, we first train a Seq2Seq- TTS system using a large amount of normal speech of one speaker and then fine-tune the learned model with normal speech of another speaker with limited data. Finally, using Lombard speech of the latter speaker, we fine-tune the model again to synthesize Lombard speech. We predict both mel-spectrograms and the World vocoder parameters as output acoustic frames. To render final speech wave- form, we employ both the WaveNet vocoder and the World vocoder.

Experiments mentioned in the model

TO experiment, the model used Blizzard Challenge 2011 speech corpus. This corpus contains utterances of 12000 voices(which add upto 16 hours). The dataset contains voices of: a) Nancy (Normal) b) Nick (Normal) c) Nick(Lombard)

We create five systems for this model experimentation The systems were different in terms of their output parameter types and the vocoder used. System S1 is the baseline system which uses a LSTM-type of recurrent neural network (RNN)- baseTTS system for adaptation, and synthesizes the speech waveform using the World vocoder. System S2 is built using the Seq2Seq-TTmodel, and the final waveform is rendered by the World vocodeSystems S3 and S4 have same architectures as systems S1 and S2respectively, but they use the WaveNet vocoder for synthesis. System S5 has the same architecture and vocoder as S4, but instead ousing the World vocoder parameters, it predicts the mel-spectrogram as the output.

Two types of listening tests were conducted: 1) speaking style sim- ilarity test and 2) comparison category rating (CCR) test of speech naturalness. The goal of the similarity test is to assess whether the technology developed is capable of generating synthetic speech of different speaking styles (normal vs. Lombard) while the CCR test aims to evaluate how much the naturalness of speech is sacrificed when the speaking style is adapted. We used an evaluation setup

Conclusion

This paper compared different TTS models and vocoders to adapt the speaking style of speech synthesis from normal to Lombard. The study proposes using an adaptation method based on fine-tuning combined with sequence-to-sequence based TTS models and the WaveNet vocoder conditioned using mel-spectrograms. Listening tests show that the proposed method outperformed the previous best method that was developed using a LSTM-RNN based adapted sys- tem. Future work includes an extensive subjective evaluations and training both the WaveNet and Seq2Seq-TTS model in a single pipeline.