34 extensions surveyed, 4 Chrome-only exclusions documented, and an answer to "is NaturalReader worth $20/month?" (Still no.) Updated 2026-05-08 with full re-verification.
\frac{x^2}{2} as "backslash frac open brace x caret two..." -- useless.
FIX MathJax's built-in accessibility (right-click equation > Speech > Auto Voicing) works on MathJax sites with zero setup; MathJax 4.0 (2026) reworked screen-reader support. For full articles: Smart TTS Reader converts LaTeX to spoken math via AI (note: dormant since May 2025, only 3 users); a more reliable DIY route is direct LLM prompting (paste LaTeX into Claude/GPT with "convert to ClearSpeak", then TTS the output). Paper2Audio summarizes equations in plain English. NotebookLM Audio Overview paraphrases math natively as part of its podcast format. Note: KaTeX still has zero accessibility/speech support as of 2026.
reader.parse-on-load.force-enabled to true in about:config. Or prefix any URL with about:reader?url=. Read Aloud, Clearly Reader, and the new Lexora (offline neural) all have their own content parsers that bypass Reader Mode entirely. macOS Tahoe 26 also ships an "Accessibility Reader" that works on selected text.
| Extension | Rating | Users | Updated | Price | TTS Engine | Outside Reader? | Selected Text? | LaTeX? | Notes |
|---|
The power user's TTS extension. Only Firefox extension connecting to multiple premium cloud TTS engines while remaining free. 217K+ users, actively maintained, open source (GPL v3). Latest v1.81.1 (Feb 2026).
Only actively maintained Firefox extension combining custom reader mode + TTS + LaTeX math rendering in one package.
The ONLY Firefox extension that explicitly converts LaTeX/MathJax to spoken math using AI preprocessing.
\sum_{i=1}^n x_i^2 becomes "sum from i equals one to n of x sub i squared"Purpose-built for academic papers, intelligently handles math by summarizing rather than reading symbols.
Purpose-built for academic papers with Firefox extension support. The most affordable paid option.
Highest-traction new TTS extension since the original report (296 users, 4.1/5). Embodies the "bring your own endpoint" pattern that became dominant in 2026 -- instead of bundling specific cloud providers, it's a UI over any OpenAI-format TTS API.
Highest-traction TTS-adjacent new extension (368 users, 5/5, updated 2026-05-08). Not a TTS engine itself but the practical Firefox bridge to Google's NotebookLM Audio Overview, which has displaced traditional TTS for many academic-paper workflows in 2025-2026.
NotebookLM produces ~20-minute conversational podcasts between two AI hosts that summarize and discuss your sources. For digesting math-heavy papers, this often beats sequential TTS reading because the AI hosts paraphrase equations into spoken English natively. NotebookLM in 2026 added interactive mode -- you can ask the AI hosts questions mid-podcast.
One of the few Firefox extensions running a fully offline neural TTS model directly in-browser (~18MB shipped). Strongest privacy story of any current extension -- no cloud round-trips, no API keys, no telemetry.
Direct competitor to Listening.com, specifically targeted at PhD students and researchers (40,000+ reported users). Adds Zotero sync -- a feature Listening.com lacks.
The Core Problem: Most TTS engines treat LaTeX as text. \frac{x^2}{2} becomes "backslash frac open brace x caret two close brace open brace two close brace" -- completely useless.
On any website using MathJax (arXiv, Stack Exchange, many academic sites):
Supports: ClearSpeak (default) and MathSpeak rule sets, 12+ languages (English, French, German, Hindi, Spanish, Korean, etc.)
MathJax 4.0 (2026): Reworked screen-reader support to use ARIA attributes from the explorer/speech components. SSML now used consistently across all rule sets.
Limitations: Only works on MathJax sites (NOT KaTeX -- still no speech support as of 2026 -- and not static images). Must interact with each equation individually. Uses basic browser voices.
Uses Gemini/Ollama AI to preprocess LaTeX into spoken English before TTS.
Example: \sum_{i=1}^n x_i^2 → "sum from i equals one to n of x sub i squared"
Requires: Gemini API key OR local Ollama + external TTS server (e.g., Kokoro-FastAPI)
Status (May 2026): Dormant since May 2025. Only 3 users. GitHub repo has 10 commits, 2 stars, 0 forks. Verify it still works before relying on it.
Best for: Power users comfortable running local AI services -- but most users will be better served by direct LLM prompting (Tier 3).
Smart TTS Reader's pipeline can be replicated in any chat UI. With the dormancy of the original extension, this is now arguably the more reliable path:
The MathReader paper (arXiv:2501.07088, ICASSP 2025) released a fine-tuned T5-small specifically as a LaTeX→spoken-English translator (WER reduced from 0.510 Edge / 0.617 Acrobat → 0.281). The model and training recipe are public; could be reused as a local, open-source alternative to Gemini.
Summarizes equations in plain language rather than reading symbols. Good for getting the gist of papers, not for precise math. Updated April 2026 (v1.8.5, 250 users).
Example: Instead of reading a complex integral, says "this equation represents the probability distribution over..."
Best for: Casual academic paper consumption where understanding > precision.
Has displaced traditional TTS for academic papers in 2025-2026. Produces ~20-min conversational podcasts between two AI hosts that paraphrase math/citations natively in spoken English -- handles equations more gracefully than literal TTS.
Pair with the NotebookLM Tools Firefox extension (368 users, updated 2026-05-08) for bulk source import (URLs, YouTube, RSS, browser tabs) and MP3 download.
2026 features: Video Overviews, voice-interactive Audio Overviews (ask AI hosts questions mid-podcast).
KaTeX accessibility status: KaTeX still produces hidden MathML for screen readers but NO built-in speech engine, no Auto Voicing, no ClearSpeak/MathSpeak. Open issues #38, #457, #820 unresolved as of 2026.
Option A: about:config → set reader.parse-on-load.force-enabled to true
Option B: Prefix URL: about:reader?url=https://...
Option C: "Activate Reader View" extension (7,549 users, 4.2/5; long unmaintained but works)
Option D: Clearly Reader and the new Lexora have their own parsers independent of Firefox.
UPDATED: macOS 26 "Tahoe" (released Sep 15, 2025) renamed "Spoken Content" to "Read & Speak".
System Settings > Accessibility > Read & Speak > "Speak Selection"
Shortcut: Option + Esc on any selected text. Works system-wide in Firefox.
Premium voices: System Settings > Accessibility > Read & Speak > System Voice > Manage Voices. (Original report incorrectly placed this under VoiceOver.) New Pan-Indian Gujarati & Marathi voices in Tahoe.
Tahoe also adds an Accessibility Reader -- displays selected text in a reader window. Personal Voice now requires only 10 phrases (down from 150) but cannot be used via Firefox's Web Speech API.
Speechify (macOS app): $139/yr, 1,000+ voices, 60+ languages (was reported 200+), works with any app including Firefox. Won 2025 Apple Design Award (Inclusivity category).
Audiblez (open source, MIT, 6.4k stars): Kokoro-based EPUB-to-audio converter. EPUB only -- not PDF/arXiv yet. — NEW
ResearchPod (iOS+web): Upload PDF or browse arXiv → ~20-min two-host podcast. 3 free generations/day. — NEW
ekoAcademic: Interactive arXiv-to-podcast with GPT-realtime; ask questions mid-listen. — NEW
Piper: Original repo archived Oct 2025; active fork at github.com/OHF-Voice/piper1-gpl. Latest 1.4.2 (April 2026). Used by Read Aloud.
Kokoro-82M: v1.0 (Jan 2025), 54 voices, 9.5M+ HF downloads/month. Powers 5+ Firefox extensions and Audiblez. Apache 2.0.
Supertonic 3 (Apr 29, 2026): On-device ONNX, 31 languages, 167× real-time on M4 Pro. Available in Read Aloud since v1.81.0.
Chatterbox (Resemble AI, MIT): Reportedly outperforms ElevenLabs in blind eval. Multilingual + Turbo variants. — NEW
F5-TTS, Sesame CSM-1B, MLX-Audio: All released/updated 2025-2026. None have native Firefox extensions yet.
As of May 2026, that's now literally the Plus plan: $20.90/mo or $119/yr (up from $19/$110 in March). NaturalReader also discontinued the entry-tier "Premium $9.99/mo" plan and added a new Pro tier ($25.90/mo or $159/yr) with HD voices and Reading Styles. Short answer: still no. Read Aloud + Google Cloud WaveNet free tier gives comparable voice quality at $0 with 4M chars/month perpetual (~64 hrs). For academic papers, NotebookLM Audio Overview is now the strongest free option. If you want to pay, Listening.com at $39/year is 67% cheaper than Plus and purpose-built for papers.
| Engine | Quality | Cost | Offline? | Notes |
|---|---|---|---|---|
| ElevenLabs | ★★★★★ | $60/yr (Starter) | No | Best-in-class; via Hearvox extension (BYO key) or Custom TTS Reader. Free tier ~10K chars/mo. |
| Chatterbox (Resemble AI) | ★★★★★ | Free (MIT) | Yes | Reportedly beats ElevenLabs in blind eval; Multilingual + Turbo variants. |
| Google WaveNet / Chirp 3 HD | ★★★★☆ | Free tier | No | 4M chars/mo perpetual via Read Aloud (was undercounted as 1M) |
| Amazon Polly Neural | ★★★★☆ | 12mo or $200 credit | No | Free tier overhauled Jul 2025 -- new accounts get $200 credit / 6mo window |
| NaturalReader Pro | ★★★★☆ | $159/yr | No | HD voices + Reading Styles (tone/emotion). NEW tier 2026. |
| NaturalReader Plus | ★★★★☆ | $119/yr (was $110) | No | Marginally better than WaveNet; web app only |
| Speechify | ★★★★☆ | $139/yr | No | 1,000+ voices, 60+ langs, macOS app only. ADA 2025 winner (Inclusivity). |
| Microsoft Azure Neural HD | ★★★☆☆ | F0 free / $30/M | No | NEW expressive tier; F0 = 0.5M chars/mo perpetual |
| Microsoft AI TTS | ★★★☆☆ | Free (Clearly Reader) | No | Used by Clearly Reader, MS Edge TTS, Narravo |
| Kokoro-82M (offline) | ★★★☆☆ | Free (Apache 2.0) | Yes | v1.0 (Jan 2025), 54 voices, 8 langs. Hit #1 on TTS Arena Jan 2026. Powers 5+ Firefox extensions. |
| Supertonic 3 (on-device) | ★★★☆☆ | Free (MIT/OpenRAIL) | Yes | v3 Apr 29, 2026: 31 langs, 167x real-time on M4 Pro. In Read Aloud since v1.81.0. |
| Piper (local neural) | ★★★☆☆ | Free | Yes | Repo moved to OHF-Voice/piper1-gpl. v1.4.2 (Apr 2026). Via Read Aloud. |
| macOS Premium Voices | ★★★☆☆ | Free | Yes | Tahoe 26: Read & Speak panel; 1-3GB per voice |
| Browser Web Speech API | ★★☆☆☆ | Free | Yes | Varies by OS/browser, often robotic |