No filming:An AI avatar or clone yourself
    Guides

    Hume Octave 2: API Ended 13 Nov 2026, What to Use

    Hume's notice ended Octave 2 API access on 13 November 2026, with account data deleted after it. What the model did and which voices replace it.

    Versely Team••9 min read

    Hume Octave 2 is Hume AI's emotional text-to-speech model, announced on 1 October 2025, and its API has reached its end date. Hume's notice set access to its TTS and EVI APIs to end on 13 November 2026 at 12:01 a.m. EST, with account data permanently deleted after that. Gemini 3.8 Flash TTS tops Hume's own voice leaderboard.

    At launch Hume listed 11 languages, audio in under 200 ms, voice cloning from a 15-second sample and half the price of Octave 1. This page covers what Octave 2 did, what carries over if you used it, and which voices replace it.

    The short answer

    • Status: Hume's developer changelog carried this entry, dated 2 October 2026, when we read it on 8 October: "Hume is sunsetting the TTS and EVI APIs. Access ends November 13, 2026 at 12:01 a.m. EST, and both APIs remain fully supported until then. Any data in your account is permanently deleted after November 13, 2026." Developers Digest quoted the same wording on 20 October, and its page still carried it on 13 November 2026. That end of access has now passed.
    • Hume has moved on: On 13 November 2026 the old changelog and text-to-speech docs addresses still redirect to Hume's Audio Decisions docs, a product launched on 2 November, and the docs index lists only Audio Decisions and Expression Measurement. Hume's homepage no longer mentions Octave. We did not test whether the API still answers.
    • Starting a new project: build on another voice.
    • If you used it: the notice says account data is permanently deleted after 13 November 2026, so the audio files and voice samples you saved yourself are what carry over. The notice also pointed to Hume's Real World VoiceEQ benchmark for alternatives.
    • Octave 2 is not in Versely's catalog. The closest voices that are, with credits per 1,000 characters, are in the table below.

    What Octave 2 is

    These details come from Hume's Octave 2 launch post (1 October 2025, still live on 13 November 2026) and its text-to-speech docs, which we read on 8 October 2026 before that address began redirecting.

    Detail What Hume says
    Announced 1 October 2025; available in the API from 3 October 2025 by sending version: 2
    Languages 11: Arabic, English, French, German, Hindi, Italian, Japanese, Korean, Portuguese, Russian, Spanish
    Speed "Under 200ms" in the launch post, and 40% faster than Octave 1
    Voice cloning From a 15-second recording of a native speaker
    Voice design Available, English only when we read the docs
    Price "Half the price of Octave 1". No flat per-character rate was published; the docs said billing was "defined by your subscription tier"
    Text limit 5,000 characters per utterance

    The launch post also said voice conversion and phoneme editing were coming soon. We could not confirm that either shipped before the sunset notice.

    What "emotional" means in Octave 2

    Octave 2 read the feeling from the script. The launch post says the model "more deeply understands the emotional tone of speech" and infers "when to whisper secrets, shout triumphantly, or calmly explain a fact". The words you wrote were the control surface.

    Plain-English direction was a separate field. Hume's acting instructions docs, as read on 8 October 2026, offered a description field, for example "whispering, hushed", and said it "is available for Octave 1 only", with Octave 2 support "coming soon". So "I type 'sound relieved' and Octave 2 obeys" was never documented.

    Most other engines handle emotion with inline tags ([whispers], [sighs]) or a preset parameter such as "excited". Natural-language direction is more flexible than a menu: a description can say "trying not to cry while staying professional", which no single preset covers. If that is the feature you want, it exists on the voices in the table further down.

    When an emotional voice is worth paying for

    Audiobook and long-form narrative performance. A novel with a grieving character or a memoir with a real arc is the job where a flat, technically clean read undermines the content.

    Character voice for games and interactive fiction. An NPC who needs to sound afraid or relieved at a specific moment is a different job from a narrator reading a manual. If the emotional state is gameplay information, a flat delivery is a bug.

    Sensitive support and companion contexts. In a mental-wellness app or a bereavement line, tone is not decoration. Getting the register wrong can cost trust in a way a shipping-confirmation voice never will.

    When it is overkill

    Anything informational at volume. Product descriptions, IVR menus, explainers and ad reads with prices and phone numbers are neutral by design. A flat, fast, cheap voice is the correct choice, and the cheapest AI voice API in 2026 has the math at scale.

    Batch and per-unit-cost-sensitive work. Narration for 100 product videos compounds the cost of an expressive model against content nobody listens to for nuance.

    Latency-bound work. Live agents and real-time dubbing are decided by time to first audio. Test it on your own stack.

    What to use instead of Hume Octave 2

    Octave 2 is not a Versely row, and with Hume's TTS API past its end date it will not become one. The fairest starting list is the one Hume itself publishes: the top four voices on its Real World VoiceEQ text-to-speech leaderboard are all on Versely, and so is the seventh.

    Voice Credits per 1,000 characters Hume's overall score Why it is close
    Gemini 3.8 Flash TTS 2 0.92 (1st) Style direction, inline tags, 130 languages
    Gemini 3.8 Flash Lite TTS 1 0.91 (2nd) Same controls at lower cost, 101 languages
    Eleven v4 Turbo 4 0.91 (3rd) Inline audio tags, 90+ languages
    Eleven v4 7 0.90 (4th) Inline audio tags, voice cloning
    Cartesia Sonic 3.6 4 0.84 (7th) Emotion from context without tags, cloning
    Inworld TTS-2 2 0.58 (25th) Highest on vocal bursts; see below

    Scores were read on 9 November 2026: a 0-1 score on expressivity and reliability, rated by people. It is one vendor's benchmark, and Hume's blog says it tested preview versions of the Gemini 3.8 voices. Credits come from Versely's catalog on the same date.

    The overall score is not the whole story for direction. Hume's Voice Controllability Leaderboard of 17 September 2026 found that "no single model dominates every form of controllability". With inline style tags, Inworld TTS 2 scored highest on vocal bursts (4.36 out of 5), and Gemini 3.1 Flash led on delivery (4.26) and prosody (4.58). That test came before Gemini 3.8.

    If you cloned a voice on Hume, the 15-second sample you recorded is what carries over, so keep it. Cartesia Voice Clone on Versely is free to create, and speaking in the clone bills per 1,000 characters on Sonic. Only clone voices you own or have permission to use.

    For more on the field, read Gemini 3.8 Flash TTS, the best text-to-speech API roundup, what audio tags are and Cartesia Sonic vs Hume Octave 2. Fix one voice per project and vary the direction per line, the same discipline as style locking in AI text to speech. For one short hook line, a free on-device voice will do.

    For a voice file that has to match a face, generate the wav first and finish with AI lipsync only if the mouth is on screen. To re-voice a performed read into another language, use the AI dubbing tool, and for plain script to voice start in AI text to speech.

    A simple test before you commit

    Read the script out loud twice: once completely flat, once performed the way you would want an actor to read it. If the two reads would land the content roughly the same way, a cheaper, flatter voice is the right call. If the flat read changes what the content means, such as a confession, an apology or a moment of relief, the performance is doing real work and that script is worth an expressive voice. A script timer tells you whether the read fits the cut.

    FAQ

    Is Hume Octave 2 still available?

    No. On 2 October 2026 Hume announced that access to its TTS and EVI APIs, which serve Octave 2, would end on 13 November 2026 at 12:01 a.m. EST, and that time has passed. Hume says account data is permanently deleted after that date, and on 13 November its voice docs redirected to its Audio Decisions product.

    What can I use instead of Hume Octave 2?

    Start with Gemini 3.8 Flash TTS, which tops Hume's own Real World VoiceEQ text-to-speech leaderboard and costs 2 credits per 1,000 characters on Versely. Eleven v4 (7 credits) and Eleven v4 Turbo (4) follow closely, and Cartesia Sonic 3.6 (4) infers emotion from context without tags.

    What is Hume Octave 2?

    Hume AI's second-generation text-to-speech model, announced on 1 October 2025. Hume said it covers 11 languages, generates audio in under 200 ms, clones a voice from a 15-second sample and costs half what Octave 1 did.

    Does Octave 2 take plain-English emotion direction?

    Not as documented when we read Hume's docs on 8 October 2026. Octave 2 inferred delivery from the script. The description field for acting instructions was documented for Octave 1 only, with Octave 2 support listed as "coming soon".

    How much does Hume Octave 2 cost?

    Hume said half the price of Octave 1 and did not publish a flat per-character rate in the launch post. Its docs said billing was defined by your subscription tier. On 13 November 2026 Hume's pricing address redirected to its homepage.

    Is Hume Octave 2 available inside Versely?

    No. Versely's text-to-speech catalog runs other engines, and the table above lists the closest ones.

    The takeaway

    Octave 2 was built around emotional delivery, but the API that serves it reached the end date Hume announced, 13 November 2026. Move that work to a voice with documented direction, and keep a flat, cheaper voice for scripts where the feeling is not the content.

    Versely AI Weekly

    New AI models, every Thursday.

    The image, video, audio and LLM models released that week, what each is good for, and what's new on Versely.

    Weekly. Unsubscribe anytime.