I'm Michael Wilson from Yamaha Corporation's Research and Development Division. I was the lead developer for Cyber Songman and I'd like to tell you some of the Voice Bank's history and details.

VOCALOID Voice Banks are based on samples of human singers, obtained by having a singer sing lines from a recording script. I wrote a new recording script for Yamaha's Cyber Diva Voice Bank (released in February 2015) which was also used for Yamaha's Fukase English Voice Bank (released in January 2016). Even before Cyber Diva's release I knew there was room to improve the script. The recording script I made ensured that the samples we recorded had the correct sounds, but it was difficult for the singers to follow. I wanted to make a script that had the best of all worlds: short, easy to read, and still free from the labeling errors that affected some earlier English Voice Banks.

I was allowed to continue to work on the recording script after Cyber Diva's development was done in March 2014, so I reviewed every diphone and triphone we could represent in English VOCALOID and searched for words that would produce the necessary sounds on first read. Because English doesn't have a strong connection between the spelling of a word and its pronunciation, and because most English speakers haven't studied phonetic transcription so we can't rely on that to prompt the sounds we need, this is harder than it may seem! A consonant or vowel may also modify the sound of nearby consonants and vowels, sometimes drastically, so avoiding those effects while keeping the script a reasonable length and still easily readable was an interesting challenge for me.

In July of 2014 the new script was complete. In the end I had to make some compromises, but my personal testing implied that the new script was easier to read while maintaining correctness. I was ready to test it with a real singer. We chose a singer with a great voice and lots of experience, so I had high hopes for the test.

We did the test recording, and the results were positive. The singer said that the script was easy to read although still not really like anything you'd normally sing (perhaps the next script can resolve this problem). Feedback from the recording session let me tune the script a bit more.

I made a test Voice Bank and verified that the results were within my expectations. Informal listening tests with my coworkers implied that intelligibility was still high and the "singing feeling" of the Voice Bank was good. It seemed like the project was a tentative success. At the time there were no plans for additional English product recordings so we set the script aside.

Later Cyber Diva and Fukase were released and confirmed that we had raised the bar for English VOCALOID intelligibility. Product planning wanted to release an American English male Voice Bank to complement our lineup. Baba-san, who made Cyber Diva's Voice Bank, suggested that if we used the same singer we used for the recording script test then we had a good chance of making a good product, since the test results had been uncommonly good. I had been working on a few techniques and software tools to help with Voice Bank creation, and while they were still rough prototypes I felt like they were ready to use on a product. So I said that as long as we could re-record the same singer I would personally make a product-level Voice Bank. And product planning took me up on the offer.

We did another recording with the singer, then I shut myself away in a studio until development was done. There were a lot of unforeseen challenges in making the Voice Bank, including me coming down with the flu at the end of the development period, but after development, testing, and release checking the Voice Bank was ready.

I was even able to squeeze two additional phonetic symbols into the Voice Bank: [4] and [@l].

[4] is the American English tap-t sound. This is close to but different from the [d] symbol, and is mostly used in place of a syllable-ending [t] which is between two vowels. For instance, "better" would become [bh e 4] [@r]. It can also be used to soften pronunciation at the end of a phrase. Some of our third-party licensees also include this symbol in their VOCALOID4 Voice Banks, and I think it's useful for creating songs with an American English accent.

[@l] is a syllabic [l] sound (that's lower-case L). When we were working on Cyber Diva there was a lot of discussion about what the best vowel was for things like the second syllable of "apple" -- this phoneme helps fill that role. It's designed to always be followed by an [l] phoneme but it can be used on its own too with varying degrees of success. Try using it in place of a [V], [U] or [@] in unstressed syllables that end in [l] for best results. For instance, the aforementioned "apple" would become [{ p] [@l l].

[参考] 0:00〜 いままでの発音
0:04ごろ〜 新しい発音記号を手動で入力（careful -> [kh e@] [f @l l]　apple -> [{ p] [@l l]　better -> [bh e 4] [@r]　get up -> [gh e 4] [V p]）

The Voice Bank also includes a growl (GWL parameter) that can add a touch of emotion or strain when applied lightly or moderately at the start or end of a word. When pegged to the max and combined with VOCALOID's Pitch Snap Mode it adds a robotic quality to the voice that may be useful for certain kinds of electronic music.

I feel very fortunate to have been part of this project, and as always I am grateful to my Research Division colleagues for their advice and support, to the development team for their work in testing and　productizing the Voice Bank, and to product planning and marketing for their efforts in sales, promotion, naming, and visual design. There are still many unsolved problems in the field of singing synthesis, and we here at Yamaha are already hard at work on the next steps for VOCALOID. To all of the VOCALOID users, fans, and critics out there, we're listening to your comments. Your feedback is valuable so please continue to put it out there and we'll continue to read it. I'm sincerely looking forward to hearing what you think about and the music you make with Cyber Songman. Thank you for your interest and support.