[v4.1.0] Multiple speech regressions: symbols announced incorrectly in Speech output but correct in AssistiveMML

[sobhanadrink]

Issue Summary

In MathJax v4.1.0, some MathML constructs are announced incorrectly when MathJax speech output is enabled.
The same expressions are read correctly when AssistiveMML is enabled, indicating a regression or inconsistency between speech generation and Assistive MathML.

This affects screen reader users relying on MathJax speech (mjx-speech / aria-label) and results in incorrect or missing mathematical meaning.

Environment

• MathJax version: 4.1.0
• Browser: Firefix
• OS: Windows 10
• Screen reader: NVDA (also reproducible via aria-label inspection)

Steps to Reproduce:

• Load the sample HTML attached below using MathJax 4.1.0
• Enable MathJax speech output (default behavior)
• Listen to the spoken output or inspect mjx-speech aria-label
• Compare with output when AssistiveMML is enabled

Expected Behavior

• MathJax speech output should match the semantic meaning conveyed by AssistiveMML.
• Common mathematical concepts (Delta, greatest integer, hat, given, frown above) should be announced correctly.

Issue 1: Delta announced incorrectly

• Problem: Δ announced as “normal triangle”
• Expected: “Delta”
• AssistiveMML reading (correct): fraction with numerator Delta d and denominator Delta t end fraction equals …

Issue 2: Greatest Integer not announced

• Problem: White square brackets ⟦ ⟧ are not announced as “greatest integer”
• Expected: “greatest integer of negative 1.7”
• AssistiveMML reading (correct): h of negative 1.7 equals left white square bracket negative 1.7 right white square bracket equals negative 2

Issue 3: Hat not announced

• Problem: f̂ is read as just “f”
• Expected: “f hat”
• AssistiveMML reading (correct): f hat of c

Issue 4: “Given” announced as “divides”

• Problem: | announced as “divides”
• Expected: “given”
• AssistiveMML reading (correct): p of open paren b given a close paren

Issue 5: Frown above not announced

• Problem: Accent ⌢ not announced
• Expected: “frown above”
• AssistiveMML reading (correct): quantity p a with frown above

Technical details:

MathJax Version: 4.1.0
Client OS: Windows 10
Browser: Firefox
Screen reader: NVDA
Speech Mode: MathJax speech (mjx-speech)

I am using the following default MathJax configuration

and loading MathJax via

<script src="https://cdn.jsdelivr.net/npm/mathjax@4.1.0/tex-mml-chtml.js"></script>

Supporting information:

• Minimal reproducible example: Included below
• No console errors
• Issue appears limited to MathJax speech generation
• AssistiveMML consistently produces correct output

Why This Is a Bug

• AssistiveMML and speech output should be semantically consistent
• Incorrect speech output results in loss of mathematical meaning

Sample HTML (Minimal Reproducible Example)
Mathjax-4.1.0-math-symbols-speech-output.html

[dpvc]

ClearSpeak has several preferences concerning how certain symbols are spoken, including the Delta and vertical line from your items 1 and 4. You can see the various preferences using the MathJax contextual menu, in the "Speech -> Clearspeak -> Select Preferences" menu. The preference for the Delta is titled "TriangleSymbol", and the vertical line is the "VerticalLine" preference.

You can set the defaults for those using the MathJax configuration. For example,

MathJax = {
    options: {
      sre: {
        style: "TriangleSymbol_Delta:VerticalLine_Given"
      }
    }
};

would set those preferences by default (rather than "auto", which is the usual default).

For 2, it looks like delimiters like parentheses are not spoken around a single term, so that "h(-1.7)" is spoken as "h of negative 1.7" without having to speak the parentheses. Apparently that applies to other delimiters like the white brackets. There is a "Paren" preference that could be used to force them to be spoken, but that will affect all delimiters, including the ones from "h(-1.7)". I have started a conversation with the author of the speech rule engine (SRE) to see if this can be improved.

The symbol you have used in 3 is U+02C6 from the Spacing Modifier Letters block of Unicode characters. This block contains characters for the IPA, UPA, and other phonetic transcriptions, and they are called "spacing modifier letters" because they are placed between two other letters in a word, rather than above or below them like other diacritical marks. They are not generally intended for use in mathematics, and so SRE doesn't include translations for them. A better choice would be U+005E (CIRCUMFLEX ACCENT), which would be read as "hat", as you desire. Note that for characters that aren't known to SRE, the literal Unicode character is included in the string output, in hopes that the screen reader will have a string for it. That doesn't seem to be the case for you.

Similarly, for 5, the character U+2322 is not one that SRE has a translation for. Alternatives include U+2312 (ARC) and U+23DC (TOP PARENTHESIS), both of which SRE can translate.

It is also possible for you to override SRE's speech for a given element using an aria-label attribute. For example

        <math xmlns="http://www.w3.org/1998/Math/MathML">
            <mrow>
                <mover>
                <mi>f</mi>
                <mo stretchy="true" aria-label="hat" accent="true">&#x2C6;</mo>
                </mover>
                <mo>&#x2061;</mo>
                <mrow>
                <mo>(</mo>
                <mi>c</mi>
                <mo>)</mo>
                </mrow>
            </mrow>
        </math>

could be used to force U+02C6 to be voiced as "hat". Note, however, that this will be used no matter what language the user has selected, as SRE won't be able to translate your label into other languages.

[sobhanadrink]

Thank you for the detailed explanation and suggestions.

Issues 1 (Delta) and 4 (vertical line) are resolved when we set the ClearSpeak preferences via the sre.style configuration, as suggested. However, we still see the same behavior when Assistive MathML is enabled, which is required for our use case.

We need assistiveMml: true for the following reasons:

• To support copying math content via the clipboard.
• To expose Assistive MathML (with aria-hidden="true") for downstream processing and accessibility control, including selectively disabling speech output (for example, by hiding speech-related nodes via aria-hidden) for certain constructs such as mover and munder, where speech output causes usability issues.

Our current configuration is shown below for reference:

window.MathJax = {
  options: {
    enableExplorerHelp: false,
    menuOptions: {
      settings: {
        assistiveMml: true,
        braille: false,
        collapsible: false
      }
    },
    sre: {
      style: "TriangleSymbol_Delta:VerticalLine_Given"
    }
  },
  output: {
    linebreaks: {
      inline: false
    }
  },
  chtml: {
    mtextInheritFont: true
  },
  startup: {
    ready() {
      const { SpeechExplorer } = MathJax._.a11y.explorer.KeyExplorer;
      SpeechExplorer.prototype.AddEvents = function () {};
      MathJax.startup.defaultReady();
    }
  }
};

With this configuration:
Could you please advise whether there is any supported alternative that would allow:

• Assistive MathML to remain enabled and
• ClearSpeak symbol preferences (such as TriangleSymbol and VerticalLine) to be honored?

For the remaining items:
Issue 3 and Issue 5 are now producing the expected speech output after applying your suggested character changes.
For Issue 2 (delimiter speech), we understand the current limitation and will wait for further updates from the SRE discussion.

Thank you again for your help and guidance.

[dpvc]

we still see the same behavior when Assistive MathML is enabled

MathJax's speech generation (via SRE and the expression explorer) and the assistiveMML extension are two different approaches to providing support for screen readers, and are not intended to be used together. (In the upcoming v4.1.1 release, selecting one will disable the other.) Using both will likely cause screen readers to read the math twice, once from the SRE speech strings that are inserted into the displayed math, and once from the visually hidden MathML.

Note that the SRE settings I gave you above control how SRE handles the speech strings. The assistiveMML extension inserts MathML into the page, and the screen reader is responsible for converting that into speech. MathJax does not control how that occurs, so that is up to the screen reader, and how that MathML is spoken may vary from screen reader to screen reader. The screen reader itself may have preferences that can be set by the user to control how some symbols are spoken, but MathJax doesn't have access to those. That is one of the drawbacks to using assistiveMML.

We need assistiveMml: true for the following reasons:

• To support copying math content via the clipboard.
• To expose Assistive MathML (with aria-hidden="true") for downstream processing and accessibility control,

I'm afraid I don't understand these reasons, as the assistive MathML is marked with the unelectable attribute and user-select: none CSS so that it is not part of selections for copy and paste operations. Unless you are changing that, or handling copy and paste via your own code, the assistiveMML extension won't help with copying math to the clipboard.

Note that MathJax's contextual menu includes the ability to copy an expression to the clipboard as MathML, though that only works for a single expression, not a paragraph or other region.

Also, the assistive MathML is not marked as aria-hidden="true", as that would mean that screen readers would not have access to it, which would defeat its purpose. So unless you are altering the MathJax output in ways that are not shown in your configuration given above, I don't understand this comment. I'm also not sure what "accessibility control" means, here.

selectively disabling speech output (for example, by hiding speech-related nodes via aria-hidden) for certain constructs such as mover and munder, where speech output causes usability issues.

Again, I'm confused by this. your configuration doesn't show anything that does this, so I'm assuming there is other code that is affecting MathJax output that you haven't shown. What nodes are you hiding, and how does that fix the issues? What "usability issues" are these?

Could you please advise whether there is any supported alternative that would allow:

• Assistive MathML to remain enabled and
• ClearSpeak symbol preferences (such as TriangleSymbol and VerticalLine) to be honored?

The assistive MML extension and SRE speech generation are unrelated, so the SRE preferences will not affect the assistive MathML, and the two should not be used together as they are alternatives for doing the same thing (supporting screen readers). Using them together will likely lead to duplicate voicing by the screen reader (unless you are modifying MathJax's output in ways you haven't shown). If you need programmatic access to your original MathML after MathJax has typeset it, that could be handled in other ways.

[sobhanadrink]

@dpvc, Thanks for the clarification.

We understand that AssistiveMML and Speech Output are alternative accessibility approaches and should not be used together. In our implementation, we enable only one at a time — when Speech Output is enabled, AssistiveMML remains present but is marked with aria-hidden="true", so screen readers read only the aria-label from the speech node and do not produce duplicate output.

We’re aligned with the fixes you suggested for most of the reported issues. The only remaining concern is Issue 2 (greatest integer symbol announcements in Speech Output), which is not reading. Hopefully this can be addressed in a future update.

Thank you for your continued support and guidance.