State of AI voice, September 2026

Hume Octave 2 is the multilingual speech-language model people keep asking us to add. It is not in the Versely catalog, and Hume is closing the API behind it: on 2 October 2026 it announced that it is sunsetting its TTS and EVI APIs, with access ending on 13 November 2026 (Hume changelog). It is no longer a pick for new work. The rows that are: Cartesia Sonic 3.5, Inworld TTS 2, ElevenLabs Multilingual, and Gemini 3.1 Flash TTS. This is a September 2026 map of those four against Octave 2 as an explainer, plus the licensing state of Suno, which is a music problem that keeps getting filed under voice.

If the job is a wav under picture, pick a TTS row and add the voiceover. If the job is a mouth, that is a lipsync plate, not a better engine.

Four voices Versely actually serves

The catalog is not a leaderboard screenshot. It is four published text-to-audio rows with different knobs. Ranks and Elo scores below are from the mid-September snapshot and have moved since. Newer rows have arrived too: Cartesia Sonic 3.6 is the newest Cartesia row at the same 4 credits, and Gemini 3.8 Flash TTS is the newer Google row at 2.

Cartesia Sonic 3.5 is the featured Cartesia TTS, released 16 June 2026. The catalog line is multilingual, highly expressive, with emotion control, speed tuning, and voice cloning. Artificial Analysis maps it as Sonic 3.5 at rank 7 on text-to-audio in the live catalog snapshot (Elo 1203). Display price is 4 credits; billing is per 1,000 characters. Categories include text-to-audio and voice-clone. That last one is the tell: you can clone a speaker here. You still get a file, not a face. Use AI voice cloning when the identity is the asset, then lay the track under the cut.

Inworld TTS 2 is Inworld's Realtime TTS-2, still marked Research Preview, released 5 May 2026. The catalog claims 100+ languages and natural-language style steering. Rank 10 on text-to-audio (Elo 1187). Display price 2 credits, same per-1,000-character billing. There is no clone category on this row. The lever is the style sentence, not a sample. That is closer to describing the feeling in a sentence than Cartesia's emotion and speed fields are, and it is the reason to open Inworld when the brief is "sound like a tired night-shift nurse" rather than "use speaker 14, slightly faster."

Gemini 3.1 Flash TTS is Google's expressive TTS, released 15 April 2026: 30 voices, natural-language style control, inline audio tags ([sigh], [laughing], [whispering]), multilingual synthesis, and multi-speaker dialogue. Rank 5 on text-to-audio (Elo 1212) — the highest of the four Versely rows. Display price 12 credits, billed per 1,000 characters. If the script is a two-hander, this is the row that actually has a multi-speaker slot. Audio tags live here the way they live in ElevenLabs: as cues in the text, not as a global mood slider. On a model that does not parse them, the brackets get read aloud. Do not copy a tagged Gemini script onto Cartesia and expect the laugh.

ElevenLabs Multilingual is the published ElevenLabs TTS row. The catalog description is Multilingual v2; the live ranking maps the family to Eleven v3 at rank 12 (Elo 1179). Display price 8 credits, billed per 1,000 characters. Cloning and the voice-change sibling are separate rows; this one is the read. Eleven's production habit is still the one most scripts arrive in: a named voice, a tagged script, a second language of the same speaker. That is text-to-speech as most people mean it. It is not Octave 2, and it is not missing from the catalog.

The door for all four is AI text-to-speech. The language roster, for the languages that actually have named voices, sits on voiceover by language. Rank numbers move. The knobs do not: clone vs no clone, tags vs a style sentence, one speaker vs two, 30 voices vs 100 languages.

Octave 2 as a comparison, not a catalog row

Hume launched Octave 2 on 1 October 2025 as a preview of its second-generation speech-language model (Hume blog). The vendor claims: 11 languages (Arabic, English, French, German, Hindi, Italian, Japanese, Korean, Portuguese, Russian, Spanish), 40% faster than Octave 1 with generation under 200 ms, half the price of Octave 1, multi-speaker conversation, more reliable pronunciation of uncommon words and numbers, plus two features Hume called first-of-their-kind for a speech-language model — voice conversion and direct phoneme editing.

That is a real product. It is not a Versely SKU. There is no /models/octave-2. A project that is built around Hume's phoneme edits or that specific 11-language set has to run on Hume's API, and that API closes on 13 November 2026. Versely will not route it, will not bill it in credits, and will not pretend a Cartesia clone is the same object.

The useful comparison is the control surface, not a fake bake-off.

Octave 2's pitch is that a model which understands the sentence will perform it: the feeling comes from the script itself. Hume's separate description field for plain-English acting instructions is documented for Octave 1 only, with Octave 2 support listed as "coming soon" (Hume docs). Cartesia Sonic 3.5 gives you emotion and speed as fields, plus a clone. Gemini 3.1 Flash TTS gives you tags in the script and a 30-voice roster, plus two speakers in one take. Inworld TTS 2 gives you a style sentence and a 100-language claim, as a research preview. ElevenLabs gives you the production default: a named voice and a tagged script. None of those is "worse Octave." They are different forms. A product explainer does not need phoneme editing. A dubbed line that has to keep the original timing might. Voice conversion — swap the timbre, freeze the phones and the clock — is Hume's answer to that; Versely's answer is a clone or a speech-to-speech row, then a lipsync pass if a mouth is in frame.

Latency is the other slide people quote. Hume's launch post said under 200 ms; later Hume docs put Octave 2 preview nearer 100 ms time-to-first-byte, excluding the network. That matters for a live agent. It does not matter for a YouTube voiceover you will caption and schedule. Do not pick a catalog row because a keynote was fast. Pick it because the knob you need is on the form.

Language coverage is the third slide. Octave 2's 11 are a published list. Inworld TTS 2's "100+" is a catalog claim on a research-preview row — believe it for a language after you have heard the take, not before. Gemini's 30 voices are a roster, not a language count. ElevenLabs multilingual is the row you already use when the same speaker has to come back in Spanish. Cartesia is multilingual and cloneable; the clone is the reason to start there when the identity has to survive a week of scripts.

Phoneme editing is the Octave 2 feature that does not have a twin on these four rows. Hume describes it as minute adjustments to timing and pronunciation, including invented phones. If that is the job — a name the roster keeps missing, a brand word that has to land a certain way — you are on Hume, not on a Versely TTS row. Voice cloning will not spell the name for you. A tagged laugh will not either.

How to pick a read this month

Start from the artefact you have to protect.

If the speaker is a person whose identity is the channel, clone on Cartesia Sonic 3.5 (or the ElevenLabs clone path) and lock that speaker. Then generate the week's scripts as takes of the same voice. Style-lock the delivery; do not re-roll a new speaker every line and hope the audience does not notice.

If the speaker is a role, not a person — a nurse, a narrator, a second character in a dialogue — Gemini 3.1 Flash TTS is the row with 30 voices, tags, and multi-speaker dialogue. Write the tags next to the clause they change. Keep them sparse. A tag on every clause is a lurch.

If the brief is a feeling in a sentence ("tired, still professional, trying not to rush the dose") and you do not have a clone, Inworld TTS 2's style steering is the Versely-shaped version of that ask. Hear it. Research Preview means the take can change. Do not build a 90-day series on a preview row without a fallback speaker on Cartesia or ElevenLabs.

If the brief is Octave 2 specifically — phoneme edits, Hume's 11-language set, Hume's conversion — leave, and do it before Hume's API closes on 13 November 2026. That work is not here. Do not prompt Sonic 3.5 as if it were Octave. You will get a Cartesia take and a ticket about why the phoneme slider is missing.

Length is the script, not a duration chip. None of these rows expose 5s / 10s / 15s. Chunk long narration; a two-minute paragraph from a blog will flatten even when the engine is fine. Mute autoplay still needs burned-in captions. The wav is not the caption track.

Credits: plan in credits, per 1,000 characters, from the catalog row you actually run. Display prices on these four sit at 4 (Cartesia Sonic 3.5), 2 (Inworld TTS 2), 12 (Gemini 3.1 Flash TTS) and 8 (ElevenLabs Multilingual) credits per 1,000 characters. That is the planning number. It is not a per-second video rate and it is not a USD API quote from Hume's site.

When the cut shows a mouth, stop. Generate or clone the voice, lock the still, then lipsync. Laying a cheap wav under a closed mouth is the production tell. The signup is for running the row the job needs, including the ones this page named.

Suno v6 and the label fight

Voice and music keep getting filed in the same folder because both come out of a speaker. They are not the same stack, and September 2026 is a bad month to pretend they are.

Suno v6 is the current flagship, introduced on 9 September 2026 (Suno); it follows v5.5 of 26 March 2026, and Suno says it will retire its earlier models. The Versely row for sound generation is Suno Sounds V6, which lists V6, Mini, Wild, V5.5 and V5 as variants: sound generation with looping, tempo, key, and lyrics subtitle capture, 2 credits a call (V5 is 1), flat billing. It writes a bed or an effect. It does not write a talking head, and it is not a licence to a label catalog. The neighboring tool is AI music generator.

The licensing state, from the public record this year: Warner settled with Suno and with Udio. Universal settled with Udio in October 2025, with a jointly licensed platform due in 2026. Universal and Sony are still suing Suno, and filed a second suit on 18 September 2026 (Digital Music News). Suno is adding audio watermarks so generated songs are identifiable, and screening uploads with Audible Magic and Musixmatch (Billboard's AI-music timeline, NPR, 18 August 2026, Engadget on Suno watermarks).

Read that as a provenance and a rights problem, not as "AI music is banned." A watermark is how a detector finds a generate. A settlement is how two parties stopped a case. A live lawsuit is not a licence. A screening pass with Audible Magic is how an upload that matches a catalog gets blocked. None of those sentences is "this 2-credit bed is cleared for a national TV spot." If the brief is a commercial that has to survive a label, you need a licence, not a prompt. If the brief is a loop under a product orbit that you generated, Suno Sounds V6 is the row, and you still disclose synthetic media when the clip is a realistic scene going to an EU audience — the Article 50 template is the label copy, not this model page.

Do not put a Suno bed under a cloned voice and call the whole file "original music plus my voice." The voice has a consent problem if it is not yours. The bed has a catalog-match problem if the prompt leaned on a living artist. Those are two different tickets.

FAQ

Does Versely serve Hume Octave 2?

No. Octave 2 runs on Hume's platform and API, and Hume is sunsetting that API (access ends on 13 November 2026). Versely's published TTS rows in this catalog are Cartesia Sonic 3.5, Inworld TTS 2, ElevenLabs Multilingual, and Gemini 3.1 Flash TTS. Compare Octave 2 as an explainer. Do not look for it under /models.

Which Versely TTS row should I start with?

Clone a real speaker on Cartesia Sonic 3.5. Write a tagged, multi-speaker script on Gemini 3.1 Flash TTS. Steer a role in language on Inworld TTS 2 and hear the preview. Use ElevenLabs Multilingual when the production habit is a named voice and a tagged script. If you need Hume's phoneme editor, you are not starting on Versely.

Can I use Suno commercially on Versely?

Suno Sounds V6 is a 2-credit sound-generation row: beds, loops, tempo, key, lyrics capture. Paid Versely plans include commercial-use licensing for what you generate here; that is not a label deal, and it does not settle Sony's case or replace a sync licence for a song you did not generate. Watermarks and upload screening exist because matching a catalog is still a live risk. Disclose synthetic media when the law or the platform asks.