A few observations, and how do you adjust vibrato when you use the singing in the text to speech mode?

[dgomez42]

Hello!

First, thank you for creating SharpVox. I’ve really enjoyed experimenting with it, especially the singing capabilities.

I have a few questions and some feedback after spending quite a bit of time testing the engine.

1. Vibrato in TTS singing mode

When using the standard Text-to-Speech singing syntax ([PHONEME<duration,note>]), I can’t seem to change the vibrato. It sounds like it always stays around 5 Hz.

I’m trying to recreate the singing style of classic synthesizers such as DECtalk, and the parameters that have worked well in my own experiments are:

• Vibrato rate: 6.5 Hz
• Vibrato depth: 0.5% (0.005 modulation factor)

A few questions:

• Is the vibrato rate fixed in TTS singing mode?
• If not, how can I adjust the vibrato rate and depth while remaining in TTS mode?
• Which source file or function implements the TTS singing vibrato?
• Would it be possible to expose those parameters through TTS mode in the future?

2. Pronunciation

I’ve noticed some pronunciation issues in ordinary TTS.

For example, the word “use” is often pronounced with the /ʒ/ sound (like the “s” in measure) instead of the expected /z/ sound.

I wasn’t sure if this is currently a dictionary issue, a grapheme-to-phoneme rule, or something else, but I thought I’d mention it.

3. High speech rates

I’ve also been testing the engine at very high speech rates.

Around 300–320 WPM, speech begins to slur noticeably.

At around 400 WPM and above, the slurring becomes much more significant, and by 600 WPM many consonants become difficult to distinguish.

What I notice most is that consonant bursts and transitions begin to lose intelligibility.

I’m curious whether this is related to the way time scaling or post-processing is currently implemented.

From listening, it almost sounds as though the overall tempo is being compressed, while some of the internal transitions are not being adjusted proportionally.

For comparison, synthesizers such as Eloquence remain quite intelligible at very high rates because every part of the signal is shortened consistently.

Is SharpVox currently using a fixed transition length between phonetic segments, or is there another part of the architecture that limits intelligibility at higher speech rates?

If there are plans to improve high-rate intelligibility in the future, I’d be very interested in testing and providing feedback.

Thank you again for all of your work on SharpVox!

[ByteSizedFox]

Hi, thanks for leaving an issue
yeah, unless you're in klattsch mode, vibrato is currently fixed, I'm actually considering making vibrato configurable as a per-voice parameter.
and yeah, pronunciation is a bit rough right now, we are currently relying on CMU pronunciation dictionary which isn't really intended to be used like this, but I will be opening a tracker issue for people to submit pronunciation errors to help fix this.
also yes, the current transition rate could use some tuning

thanks for your recommendations!
I think i will open a tracker issue where you can submit things like pronunciation cases so we aren't spreading instances of the same problem over a several threads

[dgomez42]

Hi,

Thank you for the detailed response! I really appreciate you taking the time to explain the current state of the project.

I’m glad to hear that vibrato may eventually become configurable on a per-voice basis. I think that will open up a lot of possibilities, especially for singing synthesis and experimenting with different voice characteristics.

I also think having a central pronunciation tracker is a great idea. That should make it much easier to collect examples and refine the dictionary and pronunciation rules over time. I’d be happy to contribute pronunciation cases as I continue testing.

I’m also excited to hear that you’re considering tuning the transition rate. One area that I’ve been studying recently is how classic formant synthesizers approached high-speed intelligibility. During my research, I came across technical papers describing the architecture behind Eloquence. I found them particularly interesting because they discuss how the synthesis system is organized and how tightly synchronized parameter updates contribute to smooth articulation and intelligibility.

I’m not suggesting that SharpVox should try to become Eloquence or copy its implementation. SharpVox has its own goals and architecture. I simply thought those papers might be useful as engineering references, especially if you’re experimenting with transition timing or other aspects of the synthesis pipeline. Sometimes seeing how another system approached a similar problem can spark ideas for a completely different implementation.

If you’d be interested, I’d be happy to share those papers.

I’m really looking forward to seeing how SharpVox continues to evolve. It’s exciting to see another modern formant synthesizer under active development, and I’m happy to keep testing and providing feedback as new versions become available.

https://www.synfonicaspeech.com/about#sue-hertz

[ByteSizedFox]

references to papers would be incredibly useful, i've found some on my own but any help is appriciated, rule based speech synthesis is kind of a dying art haha. resources are either rare or gatekept

[ByteSizedFox]

just changed the vibrato defaults to 6.5Hz and close to 0.5% depth

[dgomez42]

Thank you! I really appreciate how quickly you were willing to experiment with those values. I’m looking forward to trying the new defaults in the next build.

I also had a few more observations while testing.

One feature that I think could be interesting would be adding a couple more output sample rates. In addition to the existing ones, I think 16000 Hz and 10000 Hz would be nice options. I enjoy experimenting with lower sample rates, and those would capture some of the character of classic hardware synthesizers while still providing different tradeoffs in quality and bandwidth.

While testing the lower sample rates, I also noticed something that may be worth looking into. Starting around 11025 Hz, the fricative sounds, particularly /s/ and /f/, seem to become extremely weak or disappear entirely. I wasn’t sure whether this is related to the fricative generation itself or to the resampling process, but I thought I’d mention it since it was consistently noticeable during my testing.
For an eventual 8000 Hz mode, one idea might be to band-limit the frication energy appropriately for the available bandwidth rather than letting it alias or disappear. Classic synthesizers had to deal with the same limitation, and I think it could help preserve intelligibility while still maintaining the character of an 8 kHz output.

I’m really enjoying following the project’s progress. It’s exciting to see a modern rule-based synthesizer evolving so quickly, and I’m happy to keep testing future builds.

[ByteSizedFox]

I can add some extra rates, and yeah, I'm aware of the fricative issue, I'm still tracking it down, in fact SharpVox doesnt do any resampling, it natively generates at all rates, this is a bug with the fricative generation itself

[dgomez42]

That’s really interesting, and thanks for clarifying! I had already figured SharpVox was generating natively at each output sample rate, so it’s helpful to know the fricative issue has been narrowed down to the fricative generation itself. That should make it much easier to track down.

Out of curiosity, is SharpVox based on a traditional cascade-plus-parallel synthesis architecture, or is it something different? The reason I ask is because the fricative discussion made me wonder how you’ve organized the source and filter stages internally.

I’ve been reading quite a bit about classic rule-based synthesizers recently, especially Eloquence, and one of the things I find fascinating is how different systems approached frication, aspiration, and the interaction between voiced and unvoiced sounds while still maintaining high intelligibility.

Either way, I’m really enjoying following the project’s progress. It’s exciting to see someone building a modern rule-based synthesizer, and I’m looking forward to trying the next build! Some of the current architecture actually reminds me of WinTalker/Macintalk.

[ByteSizedFox]

added 10 and 16khz rates

[ByteSizedFox]

Hello!

First, thank you for creating SharpVox. I’ve really enjoyed experimenting with it, especially the singing capabilities.

I have a few questions and some feedback after spending quite a bit of time testing the engine.

1. Vibrato in TTS singing mode

When using the standard Text-to-Speech singing syntax ([PHONEME<duration,note>]), I can’t seem to change the vibrato. It sounds like it always stays around 5 Hz.

I’m trying to recreate the singing style of classic synthesizers such as DECtalk, and the parameters that have worked well in my own experiments are:

* Vibrato rate: 6.5 Hz

* Vibrato depth: 0.5% (0.005 modulation factor)

A few questions:

* Is the vibrato rate fixed in TTS singing mode?

* If not, how can I adjust the vibrato rate and depth while remaining in TTS mode?

* Which source file or function implements the TTS singing vibrato?

* Would it be possible to expose those parameters through TTS mode in the future?


2. Pronunciation

I’ve noticed some pronunciation issues in ordinary TTS.

For example, the word “use” is often pronounced with the /ʒ/ sound (like the “s” in measure) instead of the expected /z/ sound.

I wasn’t sure if this is currently a dictionary issue, a grapheme-to-phoneme rule, or something else, but I thought I’d mention it.

3. High speech rates

I’ve also been testing the engine at very high speech rates.

Around 300–320 WPM, speech begins to slur noticeably.

At around 400 WPM and above, the slurring becomes much more significant, and by 600 WPM many consonants become difficult to distinguish.

What I notice most is that consonant bursts and transitions begin to lose intelligibility.

I’m curious whether this is related to the way time scaling or post-processing is currently implemented.

From listening, it almost sounds as though the overall tempo is being compressed, while some of the internal transitions are not being adjusted proportionally.

For comparison, synthesizers such as Eloquence remain quite intelligible at very high rates because every part of the signal is shortened consistently.

Is SharpVox currently using a fixed transition length between phonetic segments, or is there another part of the architecture that limits intelligibility at higher speech rates?

If there are plans to improve high-rate intelligibility in the future, I’d be very interested in testing and providing feedback.

Thank you again for all of your work on SharpVox!

would you be willing to provide a sentence that struggles with high speed intelligibility for you? if you can do so, i can provide a sample of a test version, this would help improve intelligibility

[dgomez42]

Absolutely! I’m actually putting together a comparison now.

My plan is to generate the exact same sentences with both SharpVox and Eloquence using the same target speech rates so I can compare how intelligibility changes as the speaking rate increases. That way I’ll be listening to the same text under the same conditions instead of relying on memory.

For the Eloquence reference, I’ll be using the 11025 Hz output so the comparison stays consistent. I’m planning to test a range of speech rates, including some of the higher ones such as 300, 400, 500, and 600 words per minute, since that’s where I usually begin to notice the biggest differences in intelligibility.

Once everything is ready, I’ll upload the WAV files for both synthesizers so you can hear the comparison directly. I’ll also include my observations on where I think SharpVox begins to lose intelligibility and any patterns I notice. Hopefully that will provide some useful reference points while you’re experimenting with improvements to the high-speed speech.

[ByteSizedFox]

Ah, I just meant an example of text that SharpVox specifically struggles with.

Because I have an experimental version, but I need some inout that reliably struggles to be understood quickly so that I can actually test my new version versus the previous version. I'll definitely need some assistance on your end as a listener, the expert listener issue makes this hard for me to gauge, but comparisons would also help

[dgomez42]

Here are some sentences I just tried through the Web version.

Six slippery snakes silently slither through thick grass while several smaller snakes watch from the shadows. Please bring the blue plastic storage box before breakfast because the project briefing begins promptly. The telecommunications organization demonstrated remarkable interoperability between independent communication systems during the international conference. Open the accessibility preferences, select the speech synthesizer, adjust the pronunciation dictionary, and verify that the configuration has been successfully initialized.

I tested various speech rates. It appears that the higher I set the speech rate, the less smoother phoneme transitions occur, resulting in them sounding mushy. Testing, for example, 300 words per minute, and I went all the way up to 600 words per minute. Even testing the default rate, I noticed some pronunciation was off as well.