Connect with us

NEWS

The Pitch Knobs That No Longer Humanise AI Voice

ElevenLabs v4 will not read SSML pitch tags, OpenAI steers tone in English, and the old sliders now sit on cheaper cloud voices that streaming agents no longer use.

Published

on

ElevenLabs v3 and v4 do not support SSML at all, so the usual pitch-and-prosody kit cannot humanise those AI voices. Google still lets batch Chirp 3 HD jobs set speaking rate and prosody, while OpenAI steers delivery with a sentence of English instead of markup.

The how-to that tells developers to embed pitch tags still works on the cheaper cloud engines. The voices sold as companions are directed like actors, and they overshoot.

Eleven v4 Ignores the Old Pitch Tags

Jack Limebear’s October 2026 ElevenLabs guide, last updated on November 3, is blunt about the split. Older company models still honour a pause tag, and Flash v2 plus Turbo v2 will take an English phoneme tag. Eleven v3 and Eleven v4 read none of that markup.

They take Audio Tags written in brackets, plus whole-clip voice settings for speed and stability. The <prosody> tag that raises pitch by a few semitones is ignored. The replacement on v3 and v4 is a direction such as [slowly] or [rushed].

That is the joke hiding in the humanise-the-voice playbook. SSML is precise enough to audit. A tag that says [excited] is a stage direction, and two takes of the same line will not match.

Google’s own Chirp 3 HD scripting tips make the same move without calling it a slider. The docs contrast a stiff prompt, “The product is now available. We have new features. It is very exciting,” with a rewrite that plants hesitation in the text: “The product is now available… and we’ve added some exciting new features. It’s, well, it’s very exciting.”

The second version is still a script. It is not a pitch number. Naturalness is being pushed into the words, because the newest models would rather interpret a scene than accept a semitone.

Three Control Languages, One Job

Anyone trying to humanise AI voice in production now has to pick a dialect. Amazon Polly and Azure Speech still treat SSML as the main steering wheel. ElevenLabs v4 treats a bracketed aside as the wheel. OpenAI’s remaining Speech endpoint treats a free-text instruction as the wheel, and that endpoint is already on a shutdown clock.

HOW THE ENGINES STEER PITCH AND PACE

Engine How you steer it SSML Pitch and rate
Google Chirp 3 HD, batch SSML plus AudioConfig Yes on sync jobs speaking_rate 0.25 to 2.0; prosody tag listed
Google Chirp 3 HD, stream Plain text chunks No SSML dropped on the live path
Azure Neural SSML, list $15 per million characters Full set Pitch, rate, emphasis
Amazon Polly SSML subset of the W3C spec Yes Pauses, volume, rate, pronunciation
ElevenLabs v3 and v4 Audio Tags, speed, stability No [slowly], [rushed], clip-level speed
OpenAI gpt-4o-mini-tts English instructions No Tone, speed, whisper, accent in prose

Google Standard sits at $4 per million characters. ElevenLabs v4 Turbo and Flash sit near $0.04 per 1,000 characters, about $40 per million, roughly ten times the Google Standard rate. Azure’s neural list price, checked against Microsoft’s retail figures on October 17, 2026, is $15 per million. OpenAI’s tts-1 rate is also $15 per million.

The expensive engines are the ones that stopped exposing a clean pitch field. Teams that still need a confirmation number read back the same way every night stay on Polly and Azure. Teams that want a companion stay on tags and instructions, then pay for retakes when the warmth lands in the wrong scene.

What SSML Still Does on Cloud Engines

SSML is still the W3C Speech Synthesis Markup Language that cloud IVR stacks were built on. Amazon Polly’s developer guide says you wrap the line in a speak tag, then add pauses, volume and rate so a prompt does not gallop. Azure keeps the widest tag set of the big three clouds, including emphasis on neural voices.

Google’s Chirp 3 HD documentation lists speak, say-as, paragraph, sentence, phoneme, sub, break, audio and prosody for synchronous requests. Unlisted tags are ignored. The same page says those tags are not currently supported for streaming requests.

That matters more than a blog post about chatbot warmth. Conversational agents live on the stream. The batch path, where a developer can still write pitch in markup, is the path for files, not for a live assistant.

Google’s AudioConfig still accepts speaking_rate from 0.25 to 2.0 and pitch from -20.0 to 20.0 semitones. Those fields are numbers a QA script can check. An instruction that says “be empathetic” is not.

ElevenLabs warns that stacking too many of the old tags on the models that still accept them can yank speaking speed around and leave artefacts. The company’s own markup post says to keep to a small handful. Over-tuning is not a fan complaint. It is in the vendor note.

Bracketed Stage Directions Replace Semitones

On September 28, 2026, ElevenLabs launched Eleven v4 and Eleven v4 Turbo as its fastest and most emotive models, and said Artificial Analysis had ranked them first. v4 Turbo is also in ElevenAgents at a median inference latency of about 100 ms. The docs cap a generation at 10,000 characters and list 90+ languages.

The control surface is a line in the script, not a slider.

AUDIO TAGS THAT STAND IN FOR PROSODY

  • [whispering]: Drops the line into a hush instead of a pitch cut.
  • [sighs]: Inserts a breath the old break tag never quite faked.
  • [slowly] or [drawn out]: Replaces a rate attribute on a prosody tag.
  • [rushed]: Pushes pace up without a percentage.
  • [shouts]: Stands in for emphasis markup.
  • [excited] or [sorrowful]: Sets a mood the model will not confine to one word.
  • [pause]: The leftover cousin of a timed break, with no millisecond field.

The product now auto-tags multi-speaker takes, which is handy until the default delivery arrives too warm. A cocky or flat character comes out reciting. That is over-tuning with the new knobs: the model is doing too much acting because the tag vocabulary rewards feeling.

Price is the other constraint. Long-form generators already treat ElevenLabs as the line item that blows the budget, and cheaper engines are close enough on a first listen that people switch for volume work. Quality still wins the demo. The invoice wins the second hour.

The company posted the v4 models itself.

$22 Million for Licensed Clones

The people with money on the table are not chatbot interns. ElevenLabs said Voice Library creators had earned $11 million by November 2025 and over $22 million six months later, across 10,400+ creators. The library spans 32 languages. Only a Professional Voice Clone can be listed, and only paying-customer usage pays the owner.

Independent authors, the company said, can finish a full-length audiobook on marketplace voices for under $100 in credits. Simon Patrick, an audiobook producer and early user, has said traditional production sat at $3,000-plus a title. He split his daughter Abby’s 70,000-word manuscript into 5,000-character chunks when the first TTS model went to beta in February 2023.

Jessica Anne Bogart, a working animation voice actor, joined the marketplace under her own name.

I didn’t want to hedge or hide behind an alias. I wanted to show up as myself and give creators access to my voice, no agents, no middlemen, no gatekeeping.

Jessica Anne Bogart, voice actor, ElevenLabs Voice Marketplace

She has said marketplace income passed what she earned from the previous five years of acting. Other working voices will not go near a clone. Stacy Rivera, a Los Angeles actor who works commercials and Netflix, has said she is afraid the job will not be an option in five years. Sir Michael Caine licensed his voice to ElevenLabs for an Odyssey recording. The labour split is the product.

US audiobook sales reached $2.43 billion in 2025, up 9%, according to the Audio Publishers Association. AI titles were about 0.03% of that revenue, about $729,000. Willingness to try AI narration fell from 70% in 2025 to 61% in 2026 in the Association’s figures. Audible’s ACX path still wants a human narrator unless a title is otherwise authorised. Stores now stamp Virtual Voice, digital voice or auto-narrated on the listings that do ship synthetic files.

Why a Calmer Clone Can Sound More Trustworthy

The stated point of a natural voice is that a user feels they are with a helpful companion instead of a machine. That is also how you get people to tell a bot things they would not type.

A September 2025 PLOS One paper by Nadine Lavan, Michael Irvine, Victor Rosi and Carolyn McGettigan tested generic ElevenLabs-style voices and clones against human recordings. They did not find a hyperrealism effect. They did find that generic AI voices were rated more trustworthy than the human voices, 57.9 against 53.8, and more dominant as well.

WHAT THE LISTENER STUDIES ACTUALLY SHOW

  • Trust score: Generic AI voices sat at 57.9 against 53.8 for human voices in that PLOS One pair.
  • Revenue share: AI audiobooks were 0.03% of US audiobook sales in 2025, about $729,000.
  • Willingness: Listeners ready to try AI narration dropped from 70% in 2025 to 61% in 2026 in APA research.
  • Disclosure: OpenAI’s usage policies already require a clear notice that a TTS voice is AI-generated.

A companion that sounds calm will be believed. The PLOS pair is small in absolute terms and it still points the wrong way for anyone shipping a money or health agent. The fix is not a warmer similarity boost. It is a label, and a voice that does not plead.

Background noise and fake room tone, the other trick in the humanise kit, make that label harder to hear. They also make a cheap speaker sound worse. Device testing is still the unglamorous half of the job, because a voice that passes on studio monitors will break on a kitchen smart speaker.

OpenAI Will Retire Its Speech Models in 2027

OpenAI’s Speech docs, current in mid-November 2026, say tts-1, tts-1-hd and gpt-4o-mini-tts shut down on January 6, 2027. New work is supposed to move to gpt-realtime-2.1-mini on the Realtime API, which is not a drop-in for /v1/audio/speech. Existing apps can keep the old workflow until that date.

The sample the docs still publish does not set pitch. It asks for a cheerful and positive tone. The same guide says gpt-4o-mini-tts can be prompted for accent, emotional range, intonation, impressions, speed, tone and whispering. Those instructions do not work on tts-1 or tts-1-hd. The endpoint lists 13 built-in voices and recommends marin or cedar for quality. Input is capped at 4,096 characters, a different limit from ElevenLabs’ 10,000.

THE CONTROL LAYER, IN ORDER

  1. February 2023: ElevenLabs puts an early TTS model in beta and people start feeding it novel chapters in 5,000-character slices.
  2. November 2025: Voice Library creators hit $11 million in earnings.
  3. May 22, 2026: That total is reported as more than $22 million across 10,400+ creators.
  4. September 28, 2026: Eleven v4 and v4 Turbo launch, with Audio Tags and no SSML on those models.
  5. January 6, 2027: OpenAI retires tts-1, tts-1-hd and gpt-4o-mini-tts.

The leftover craft is still real. Vary pause and pace on the engines that still accept numbers. Keep emotion tags sparse. Test the same line on the hardware the user owns. Do not ship one stability value across a legal disclaimer and a joke.

OpenAI’s usage policies require a clear disclosure to end users that the TTS voice they are hearing is AI-generated and not a human voice. The Speech API that currently carries that notice shuts down on January 6, 2027.

Harry is the editor and lead writer of CUMBERNAULD MEDIA, which he runs as an independent publication after a decade in journalism spent moving from reporting to editing. His habit is to open the document before the summary of it. A company result is read from the filing rather than the press release, a court or regulatory decision from the judgment itself, a scientific finding from the paper and its methods section rather than the headline claim, and a sporting sanction from the governing body's own ruling. That approach shapes coverage across news, business and technology as much as science, sports and entertainment, and it carries into the lifestyle, travel, auto and gaming pages, where product specifications are checked against the manufacturer's sheet and, where possible, against Harry's own testing. Every number is checked before publication, and where a source's figures are disputed the story says so. Corrections follow a public policy and are marked on the page. Readers anywhere in the world who write in get a reply from him, and the address is support@cumbernauld-media.com.

Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Trending