How we checked: Hume's shutdown notice and old prices come from Hume's own changelog and pricing page as archived by the Wayback Machine, because both live pages now redirect to Hume's new product. Every other price comes from the vendor's own pricing page, read November 11, 2026. We converted text-to-speech prices to a cost per hour of audio so they can be compared. We did not run these APIs ourselves; the only latency numbers here are vendor claims (labelled) and one independent benchmark.
What Hume announced, and what is left#
On October 2, 2026 Hume posted a "TTS and EVI API sunset notice" to its developer changelog: "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." The same notice says any data in your account is permanently deleted after that date (Hume changelog, archived October 3, 2026).
Hume the company is not closing. Its homepage now sells an Audio Decisions API, which answers questions about a recording (which emotion a caller shows, which language is spoken, how many people are talking) without a transcript. Its developer docs and changelog now cover only that product, and the old Octave documentation URLs redirect to it. So if you used Hume to measure emotion in calls, that part of Hume lives on in a different form. If you used Hume to speak, through Octave or EVI, you need a new vendor.
Two migrations, not one#
Hume sold two different things, and they have different replacements:
- EVI (Empathic Voice Interface) was a speech-to-speech API: audio in, audio out, with the language model in the loop. Its replacements are realtime voice APIs and managed voice-agent platforms, priced per minute of conversation.
- Octave was a text-to-speech API: you send text, you get speech, with control over delivery. Its replacements are TTS APIs, priced per character or per minute of audio.
EVI alternatives: realtime voice APIs compared#
For reference, Hume's last public pricing page (archived September 30, 2026) included EVI minutes in plans at $0.07 down to $0.04 a minute, with extra EVI 3 minutes at $0.06 to $0.04. Here is what the realtime alternatives charge today:
| API | How it is priced | Per minute, as published | What is included |
|---|---|---|---|
| OpenAI Realtime (gpt-realtime-2.1) | Per token | Not published per minute: $32 per 1M audio input tokens, $64 per 1M audio output tokens (mini: $10 / $20) | Speech in, model, speech out |
| Gemini 3.8 Live (Google) | Per token, with Google's per-minute estimates | $0.005 per minute of audio in, $0.018 per minute of audio out, plus text and thinking tokens | Speech in, model, speech out |
| ElevenLabs Realtime v1 Mini / Max | Per minute | $0.08 per minute plus LLM usage | Speech-to-speech model; LLM billed separately |
| Ultravox (now owned by Inworld) | Per minute | $0.05 per minute after 30 free minutes; Pro $100 a month | Speech-native model and hosted agent platform |
| Deepgram Voice Agent API | Per minute of connection | $0.075 Standard; $0.065 if you bring your own LLM; $0.050 bringing your own LLM and TTS | Speech-to-text, a managed LLM and text-to-speech in one rate |
| Rime Prism | Per minute | $0.080 (Flex); $0.064 on the $100-a-month Scale plan | Speech-to-speech model |
| Cartesia Managed Agents | Per minute | $0.03 agent fee (cut from $0.06), plus $0.014 a minute for a Cartesia phone number; LLM billed separately | Agent runtime on Cartesia's own speech models |
Which EVI replacement fits#
- You want the same shape as EVI, one socket with the model inside: OpenAI Realtime or Gemini Live. Google publishes per-minute estimates for Gemini Live ($0.005 a minute of audio in, $0.018 out); OpenAI publishes only per-token prices for its realtime models, which makes a monthly bill harder to forecast. ElevenLabs' Realtime v1 says it is OpenAI Realtime-compatible, which can shorten a port if you already target that protocol.
- You want one flat per-minute bill for a phone or web agent: Ultravox charges a flat $0.05 a minute after 30 free minutes. Inworld bought Ultravox on September 30, 2026; Inworld's announcement says existing voice IDs keep working. Deepgram's $0.075 Standard rate covers speech-to-text, a managed LLM and text-to-speech in one price, by Deepgram's own description.
- You used EVI's emotion awareness: no replacement here reads the caller's emotion the way Hume marketed EVI to. You can keep using Hume for that side through its Audio Decisions API, or have your LLM infer tone from the transcript. Treat any vendor claim of "empathic" or "emotional" voice as marketing until you hear it on your own calls.
- You would rather own the pipeline: open-source orchestrators such as Pipecat and LiveKit Agents let you swap the speech-to-text, LLM and TTS vendors yourself, so the next shutdown costs you a config change, not a rewrite. Managed platforms such as Vapi and Retell AI sit on top of these models; our voice agent platforms guide compares them.
Octave alternatives: text-to-speech APIs compared#
Octave's archived pricing charged $0.15 down to $0.05 per 1,000 characters past each plan's allowance, and the $500-a-month Business plan included 10 million characters, or $50 per million. To compare vendors that bill per character with ones that bill per minute, we converted everything to cost per hour of generated audio, assuming 800 characters per minute (48,000 characters an hour). Cartesia's own estimate is 750 to 800 characters per minute; Inworld and Rime assume about 1,000, which would make per-character vendors look about 25% dearer than shown.
| TTS API | Published price | Per hour of audio (our conversion) | Languages (vendor) | Voice cloning |
|---|---|---|---|---|
| Hume Octave (ending Nov 13) | $0.15 to $0.05 per 1K chars overage | $7.20 to $2.40 | n/a | Yes, until shutdown |
| Inworld TTS-2 Flash | $15 per 1M chars on demand; $7 on Growth | $0.72; $0.34 on Growth | 200+ (pricing page) | Instant and professional clones |
| Inworld TTS-2 | $25 per 1M chars on demand; $12.50 on Growth | $1.20; $0.60 on Growth | 200+ (pricing page) | Instant and professional clones |
| OpenAI tts-1 / tts-1-hd | $15 / $30 per 1M chars | $0.72 / $1.44 | Not compared | Not offered on the pricing page |
| Deepgram Aura-2 / Flux TTS | $30 / $45 per 1M chars (Aura-2 $27 on Growth) | $1.44 / $2.16 | Aura-2: 7 | Not offered on the pricing page |
| Rime Mist v3 / Coda v2 | $0.030 / $0.050 per minute of audio (20% off on Scale) | $1.80 / $3.00 | Mist v3: 4; Coda: 10 | Custom cloning on Enterprise only |
| Cartesia Sonic-3.6 | Plan credits, 1 credit per character: Pro $5 for 100K, Startup $49 for 1.25M, Scale $299 for 8M | $2.40 (Pro) to $1.79 (Scale) | 44 | Instant on Pro; professional from Startup |
| ElevenLabs Flash v2.5 / v4 Turbo | $0.04 per 1K chars ($40 per 1M) | $1.92 | v4 Turbo: 90+; Flash: 32 | Yes |
| ElevenLabs v4 | $0.08 per 1K chars ($80 per 1M) | $3.84 | 90+ | Yes |
| MiniMax speech-2.8 Turbo / HD | $60 / $100 per 1M chars | $2.88 / $4.80 | 40 | Rapid cloning $1.50 per voice; voice design $3 |
Latency, measured independently: Coval, which sells voice-agent testing to many of these vendors, times every model from sending text to the first audible sound. Over the 30 days to October 28, 2026 its board had Inworld TTS Flash 2 at 81 ms, Inworld TTS 2 at 166 ms, ElevenLabs v4 Turbo at 194 ms and Flash v2.5 at 198 ms, Deepgram Flux TTS at 203 ms, Cartesia Sonic 3.6 at 292 ms, Deepgram Aura 2 at 300 ms, and Rime Mist v3 at 319 ms and Coda at 339 ms (Coval TTFA benchmark). Vendor figures are lower because most stop the clock at the first byte; Coval counts leading silence too.
Inworld TTS-2: cheapest and fastest in this group#
Inworld TTS-2 Flash is the lowest per-character price on this list and the fastest of these vendors in Coval's benchmark. TTS-2 adds natural-language steering of pace and tone, which is the nearest thing here to Octave's acting instructions, and the pricing page includes voice design and cloning. Weak spots: the prices you see depend on the plan tier, so model your volume before comparing; professional clones on the free On-Demand tier need a card on file; and Inworld does not publish a per-minute price for its speech-to-speech product, so EVI users cannot price that path from the page. Stork earns a commission on Inworld sign-ups through the links on this page.
ElevenLabs: most languages and expressive control, at a higher price#
ElevenLabs v4 lists 90+ languages and audio tags for fine-grained delivery, and Flash and v4 Turbo are built for low latency (about 75 ms and 100 ms, by ElevenLabs' own figures). It also sells both a speech-to-speech API and a full agent platform (ElevenLabs Agents, $0.08 a minute plus LLM). Weak spots: at $40 to $80 per million characters it costs two to five times Inworld per hour of audio, and Agents pricing adds the LLM on top. Stork earns a commission on ElevenLabs sign-ups through links on this page.
Cartesia Sonic: the ready-made migration path#
Cartesia publishes a Hume Octave migration guide that maps Hume's request fields to its own API and keeps Hume behind a flag until you switch. That is a competitor's page, so read it as a sales tool, but it is the most concrete porting help we found. Sonic-3.6 covers 44 languages. Weak spots: pricing is plan credits, so a low-volume user pays the Pro rate of about $2.40 an hour; and Coval found that 45% of Sonic 3.5's time to first audio was leading silence. No affiliate relationship.
Deepgram, Rime, OpenAI and MiniMax#
- Deepgram Aura-2 is a good fit if you already use Deepgram for speech-to-text: $30 per million characters, seven languages, and Deepgram now recommends its newer Flux TTS ($45) for new builds. No voice cloning is listed. No affiliate relationship.
- Rime bills per minute of audio, which makes cost easy to forecast, and its free tier gives $50 of usage a month. Mist v3 covers 4 languages and Coda 10. In Coval's benchmark Coda showed a 10.7% word error rate, against 1.5% to 3.1% for every other model on the board, so test it on your own scripts. No affiliate relationship.
- OpenAI tts-1 ties Inworld Flash as the cheapest per character here, and gpt-4o-mini-tts is billed per audio token ($12 per million output tokens). No self-serve voice cloning is listed. No affiliate relationship.
- MiniMax speech-2.8 covers 40 languages and sells voice design from a text description ($3) and rapid cloning ($1.50 a voice), the closest match to Octave's voice-design feature. It is the most expensive per character on this list. No affiliate relationship.
Enjoying this? Get one like it in your inbox each morning.
one email a day · unsubscribe in two clicks · no third-party tracking
Picks by use case#
| If you used Hume for | Start with | Why |
|---|---|---|
| An EVI voice agent on the phone | Ultravox or Deepgram Voice Agent | Flat per-minute rates; Deepgram's includes the LLM |
| An EVI-style single realtime API | Gemini Live or OpenAI Realtime | Speech in and out over one connection; Google publishes per-minute estimates, OpenAI only per-token prices |
| Octave in a latency-sensitive agent | Inworld TTS-2 Flash | Lowest price per hour here and 81 ms in Coval's benchmark |
| Octave with acting direction or voice design | ElevenLabs v4, Inworld TTS-2 or MiniMax | Audio tags, natural-language steering, or voice design from a prompt |
| Octave in many languages | ElevenLabs v4 or Inworld | 90+ and 200+ languages by the vendors' own counts |
| The least porting work | Cartesia Sonic | Publishes a Hume-specific field mapping |
| Emotion detection in calls | Hume's Audio Decisions API | The part of Hume that continues |
A migration checklist#
- Today: download generated audio you need and the source recordings for every cloned voice. Note which Hume voices and settings each feature uses.
- Find the calls: search your code for api.hume.ai and your Hume API key variable. Octave calls go to /v0/tts; EVI uses a WebSocket.
- Put the new vendor behind a flag so you can switch back for a day if a voice sounds wrong. After November 13 there is nothing to switch back to, so test before then if you can.
- Re-create voices from the saved source audio on the new vendor, and listen to them on your real scripts, not demo sentences.
- Re-check pricing at your real volume. Plan tiers cut the rate by half or more at Inworld, 20% at Rime and 10% at Deepgram.
- Design for the next shutdown. PlayHT and LMNT also closed in the last year; see our list of voice APIs that shut down and what replaced them. An orchestrator such as Pipecat or LiveKit keeps the vendor a config value.
Frequently asked questions#
When does Hume's EVI and Octave TTS API shut down?#
Access ends November 13, 2026 at 12:01 a.m. EST, according to Hume's October 2, 2026 changelog notice. Both APIs are supported until then, and Hume says account data is permanently deleted after that date.
Is Hume AI shutting down completely?#
No. Hume is ending its text-to-speech (Octave) and speech-to-speech (EVI) APIs, but the company continues with a new Audio Decisions API that answers questions about audio, such as which emotion a caller shows.
What is the best replacement for Hume EVI?#
For the same single-API shape, OpenAI Realtime or Gemini Live. For a flat per-minute phone agent, Ultravox ($0.05 a minute) or Deepgram's Voice Agent API ($0.075 a minute including a managed LLM). Prices from each vendor's page on November 11, 2026.
What is the cheapest Hume Octave TTS alternative?#
Inworld TTS-2 Flash and OpenAI tts-1 both list $15 per million characters on demand, about $0.72 per hour of audio at 800 characters a minute. Inworld's Growth plan lowers Flash to $7 per million.
Can I move my Hume cloned voices to another provider?#
Not directly: voice IDs do not transfer between vendors. Keep the original recordings you used to make each clone and re-create the voice on the new provider before Hume deletes account data after November 13, 2026.
Which TTS API has the lowest latency after Hume?#
In Coval's independent benchmark (30 days to October 28, 2026), Inworld TTS Flash 2 averaged 81 ms to first audio, ElevenLabs v4 Turbo 194 ms, Deepgram Flux TTS 203 ms and Cartesia Sonic 3.6 292 ms. Coval sells testing services to voice vendors.