- Home
- Skills
- Writing & Content
- Edge TTS: Neural Text to Speech
More screenshots
Works with the AI Agents you already use
Edge TTS: Neural Text to Speech
Free neural text-to-speech via Microsoft Edge voices: casting, speech-ready text prep, SSML direction, chunked long-form rendering, subtitles.
$5
Edge TTS: Neural Text to Speech
Example session with this skill installed
Turn this changelog note into one MP3 under 30 seconds, one narrator: "On Oct. 23, 2026, Dr. Patel shipped v3.2 — approx. 35% faster search for U.S. users. See the changelog for details; the old index retires after Fri." Show me exactly what you'd feed the synthesizer and why.
- Read your context and instructions
- Compiled the edge tts
Job route: announcement — single request, full prep, light SSML. Under 2,000 characters and no captions requested, so no chunking (per the decision table).
Prep pass — six traps caught before synthesis:
- "Oct. 23, 2026" → "October twenty-third, twenty twenty-six" (written as it should be heard, not as digits)
- "Dr." → "Doctor"; "Patel" test-read in the candidate voices and kept as-is — the pinned voice handles it cleanly
- "v3.2" → "version three point two"; "approx." → "approximately"; "35%" → "thirty-five percent"
- "U.S." → "United States" to prevent a letter-by-letter read; "Fri." → "Friday"
- The em dash becomes a 400ms break — a held breath before the benefit, instead of the engine guessing at punctuation
- "retires" left alone; no homograph risk in context, noted in the job notes
What actually gets synthesized (SSML):
<speak><voice name="en-US-AriaNeural">On October twenty-third, twenty twenty-six, Doctor Patel shipped version three point two, <break time="400ms"/> with approximately thirty-five percent faster search for United States users. See the changelog for details. The old index retires after Friday.</voice></speak>
Render settings: voice en-US-AriaNeural (won the scorecard on number handling for changelog work), rate +5%, pitch +0% — pinned in the project's tts.json so the next changelog matches this one.
QC: listened end to end at 1x — no digits spelled out, no letter-by-letter reading, break lands where the dash was. Runtime about 21 seconds, inside the 30-second cap. Delivered as a single MP3; total synthesis requests: one.
Connects securely to your tools. The creator never sees your data.
About this skill
Microsoft Edge's neural voices are free, need no API key, and sound genuinely good — yet most text-to-speech output still sounds like a robot reading a spreadsheet. The gap is almost never the voice. It is the input: markdown asterisks read aloud, prices guessed wrong, names mangled differently in every chapter, and a rate knob cranked to get through an overlong script. This skill is the workflow that fixes the input side.
It starts with routing: a decision table maps the job — voice note, announcement, long narration, multi-speaker dialogue, captioned media — to its preparation depth, rendering unit, and stitching method, so short jobs stay one command and long jobs get the chunked treatment. Voice casting is done by ear with a five-criterion scorecard on a shared test paragraph, because register fit beats accent stereotypes, and the winning voice gets pinned in a small config file so re-renders months later still match.
The preparation pass is the heart of it: strip formatting, expand abbreviations and numbers the way they should be heard, split two-breath sentences, and log every pronunciation fix in a per-project glossary. SSML covers the directed-reading moves that punctuation cannot — timed breaks, sparing emphasis, and say-as or phoneme fixes for names and codes. Long documents are split at paragraph boundaries into numbered chunks, synthesized checkpoint-style (a failure costs one chunk, not the document), and stitched with ffmpeg. Word-boundary timings are captured per chunk and offset correctly, yielding subtitle-ready files instead of timings that drift against the stitched audio.
A worked podcast cold-open example, an edge-case section (mixed-language scripts, numbers-heavy financial reads, listen-in-full compliance text), and a troubleshooting table for the service's real failure modes are included.
- It is a hosted free service, not a contracted API: heavy or rapid-fire batch use can be throttled, and the chunked workflow exists partly to ride that out.
- Speech only — no singing, no voice cloning, and no real-time conversational latency.
- A voice occasionally mangles an unusual name beyond what prep can fix; the glossary workflow catches repeat offenders, but a first render of name-dense text needs a full listen.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 24 days ago