Abstract
We propose an end-to-end speech synthesizer, Fast DCTTS, that synthesizes speech in real time on a single CPU thread. The proposed model is composed of a carefully-tuned light-weight network designed by applying multiple network reduction and fidelity improvement techniques. In addition, we propose a novel Group highway activation that can compromise between computational efficiency and the regularization effect of the gating mechanism. As well, we introduce a new metric called Elastic Mel Cepstral Distortion(EMCD) to measure the fidelity of the output mel-spectrogram. In experiments, we analyze the effect of the acceleration techniques on speed and speech quality. Our best model maintains a speech quality similar to that of the baseline model, DCTTS, with the computation reduced to 1.76% and the number of parameters decreased to 2.75%. The speed on a single CPU thread was improved by 7.45 times, which is fast enough to produce mel-spectrogram in real time without GPU.FastDCTTS vs. baseline DCTTS - LJSpeech dataset
| No. | baseline 90% | FastDCTTS 90% | baseline 70% | FastDCTTS 70% | Script |
|---|---|---|---|---|---|
| 01 | The most trifling acts were magnified into offenses. | ||||
| 02 | One smug alderman, a member of the House of Commons, sneered at the ultra philanthropy of the champions of prison improvement. | ||||
| 03 | First by daily services, the latter by the appointment of schoolmasters and instruction in reading and writing. | ||||
| 04 | as is shown by the report of the Commisioners to inquire into the state of the municipal coporations in | ||||
| 05 | It was understood also that he had served in the army as a private, and had, moreover, undergone a sentence of transportation. |
FastDCTTS vs. baseline DCTTS - KSS dataset
| No. | baseline 90% | FastDCTTS 90% | baseline 70% | FastDCTTS 70% | Script |
|---|---|---|---|---|---|
| 01 | 한국은 천연자원이 풍부하지 않습니다. | ||||
| 02 | 그들은 불을 피우기 위해 나무를 모으고 있었다. | ||||
| 03 | 이 책에는 새로운 것이 하나도 없어. | ||||
| 04 | 이 자리는 임산부를 위한 자리입니다. | ||||
| 05 | 우리 담임선생님은 굉장히 엄한 분이세요. |
Weight-norm trick on baseline vs. baseline DCTTS - LJSpeech dataset
| No. | baseline | weightnorm trick | Script |
|---|---|---|---|
| 01 | The most trifling acts were magnified into offenses. | ||
| 02 | One smug alderman, a member of the House of Commons, sneered at the ultra philanthropy of the champions of prison improvement. | ||
| 03 | It was understood also that he had served in the army as a private, and had, moreover, undergone a sentence of transportation. |
Weight-norm trick on baseline vs. baseline DCTTS - kss dataset
| No. | baseline | weightnorm trick | Script |
|---|---|---|---|
| 01 | 어제 주유소를 들러서, 기름을 가득 채웠습니다. | ||
| 02 | 눈에 뭔가가 들어갔어요. | ||
| 03 | 서점에서 아이디어는 좀 얻었어? |
Visualization of EMCD alignment
