NVDA ALPHA AND 2025.3beta3 SAPI4 synthesizer volume becomes abnormally loud during startup when navigating log viewer

[dpy013]

    ### Steps to reproduce:

First, download the latest NVDA alpha version and 2025.3beta3.
Then, create a portable version, launch it, and switch the speech synthesizer to SAPI4.
Finally, open the NVDA menu, go to Tools > Log Viewer, position the arrow keys on the line "Starting NVDA version alpha-52411,952db62f x86," and use the left and right arrow keys to browse—this will reproduce the error.

Actual behavior:

The voice output of SAPI4 is abnormal.

Expected behavior:

The voice output of SAPI4 should be normal.

NVDA logs, crash dumps and other attachments:

nvda.log

System configuration

NVDA installed/portable/running from source:

NVDA version:

alpha-52411,952db62f

Windows version:

version 24H2 (OS 26100.5074)

Name and version of other software in use when reproducing the issue:

The issue primarily occurs with the vvtts component of SAPI4.

Other information about your system:

none

Other questions

Does the issue still occur after restarting your computer?

yes

Have you tried any other versions of NVDA? If so, please report their behaviors.

none

If NVDA add-ons are disabled, is your problem still occurring?

yes

Does the issue still occur after you run the COM Registration Fixing Tool in NVDA's tools menu?

yes
Below is a specific screen recording demonstration:

2025-09-05_07-00-26.mp4

[SaschaCowley]

I couldn't reproduce this issue with Microsoft Mary, Mike or Sam, which are the only SAPI4 voices I have access to.

• Can you try with some prior versions of NVDA and see if you can still reproduce the issue?
• What voices have you tried this with? Some voices have strange volume anomalies when reading by character.
• Can you reproduce this when reading other content by character, or is it only in the NVDA log viewer?

[dpy013]

I tested both nvda_2025.3beta3 and nvda_snapshot_alpha-52402, and both encountered the same issue mentioned above.
Additionally, I switched the language of NVDA to English.
Here is an IBM ViaVoice TTS speech synthesizer to see if you can reproduce the issue?

[dpy013]

After further testing, it was found that when using uppercase and lowercase input of English letters in Windows 11 Notepad, the sound also suddenly becomes louder. Below is a screen recording demonstration:

2025-09-08_07-45-36.mp4

[seanbudd]

cc @gexgd0419 - could this be a regression from your recent work?

[dpy013]

Additionally, 2025.3 beta3 also has this issue.

[seanbudd]

Did 2025.2 have this issue?

[dpy013]

I tested again with 2025.2 on my other computer and was able to reproduce the issue mentioned above. Below are the logs from Notepad and the NVDA log viewer. I cannot record a screen video at the moment because the environment at my workplace is too noisy, and my office laptop does not have an external sound card:

nvda.log

notepad.log

[seanbudd]

@dpy013 - do you know which is the most recent version of NVDA which works correctly? does 2025.1 work fine? 2024.4?

[dpy013]

Tested with 2024.4 and encountered the same issue mentioned above. I might need to rule out the problem with NVDA. Below is the problematic speech TTS:

IBMViaVoiceTTSRuntime.exe.gz

[gexgd0419]

It seems that the SAPI4 interface of VVTTS does not handle volume level greater than 65535 properly in a Vol tag.

According to SAPI4's spec, when the volume is greater than 65535, it should be separated into two words, where the low word is for the left channel and the high word is for the right channel. The current SAPI4 code set both channels to the same volume by duplicating the low word to the high word, which should be accepted by SAPI4 voices according to the spec.

The SAPI4 interface of VVTTS calls ibmeci.dll to actually synthesize the speech, so all SAPI4 parameters will be converted to a format that eciSetVoiceParam can accept.

When calling ITTSAttributes::VolumeSet to set the volume, it is handled correctly. When the low word and high word are the same, only the low word will be passed to eciSetVoiceParam. When the two words are different, their average value will be put in the low word and passed to eciSetVoiceParam. So it seems that VVTTS does not support different volume level in different channels, and eciSetVoiceParam only accepts 0~0xFFFF as a valid volume level.

When using a Vol tag in the text to set the volume, however, the volume level is not checked or processed; instead it is directly sent to eciSetVoiceParam, which only expects 0~0xFFFF. Therefore, a volume level with both words set is "invalid", and will be ignored by eciSetVoiceParam.

Quote from the Microsoft SAPI 4 docs for Vol tags:

The volume level is a linear range from 0 for absolute silence to 65535 for the maximum monaural volume. The default is 65535.

If you specify a value greater than 65535, the engine assumes that you want to set the left and right channels separately and converts the value to a double word, using the low word for the left channel and the high word for the right channel. For example, a value of 65536 sets the left channel to the maximum baseline speaking volume and the right channel to the minimum.

The SAPI4 implementation of VVTTS satisfies the first paragraph, but not the second.

Changing Vol tags to specify only the low word works for this voice. Theoretically, this could work for other SAPI4 voices as well, assuming that they all follow the SAPI4 spec and correctly recognize volume levels between 0~0xFFFF as same volume for both channels. But we have already seen a few instances where some voices don't follow the SAPI4 spec, so ... this would need some thorough tests at least.