Hume is shutting down its text-to-speech and EVI APIs. Access ends on November 13, 2026 at 12:01 a.m. EST, and Hume says any data left in your account is permanently deleted after that (Hume’s notice, posted October 2, 2026). If your product calls Octave or EVI, you need a replacement live before that date.
Hume itself is not closing. It now sells VoiceEQ, a platform for evaluating voice AI, along with expression measurement APIs. So if you used Hume to measure vocal expression rather than to generate speech, check whether the Expression APIs still cover your case before you migrate anything.
What to move before November 13
Your audio and transcripts. Download any generated audio, EVI chat transcripts, and logs you need from your Hume account. After November 13 they are deleted.
Your voices. Voices you designed with a prompt or cloned on Octave live on Hume’s servers and do not export. Keep the original recordings you used for a clone; with them, you can recreate the voice on another provider. A designed voice has no source recording, so pick its closest match from another provider’s library and test it on your real scripts.
Your integration. Hume voice IDs and EVI configs do not carry over. Budget time to map each config’s voice, prompt, and tools onto the new provider, and to retest turn-taking and interruptions if you ran EVI on phone calls.
What to compare
Octave and EVI did different jobs, so they need different replacements. Octave was text to speech: a TTS API replaces it. EVI was a full speech-to-speech voice interface: replacing it takes speech recognition, a language model, and speech output, either as separate APIs or as one agent platform. Cartesia’s Sonic generates speech, Ink transcribes speech, and Managed Agents builds voice agents from both. Expressive output is not evidence that a system can identify a person’s emotional state.
The options below cover different jobs; they are not a ranked benchmark. Check each provider’s current availability, features, and plan terms before committing.
Alternatives at a glance
| Product | Evaluate for | Check before choosing |
|---|---|---|
| Cartesia | Speech for voice applications | Model, language, and integration requirements |
| OpenAI | Speech and conversational APIs | Endpoint choice and conversation control |
| IBM Watson Text to Speech | TTS API integration | Languages, controls, and service limits |
| Microsoft Azure Text-to-Speech | Speech in Azure applications | Region, SSML support, and custom voice access |
| Amazon Polly | TTS in AWS applications | Voice engine, region, and supported controls |
| Speechify | Reading documents aloud | Reading app, studio, or API plan |
| Murf AI | Scripted voiceovers | Editing controls and export rights |
| Descript | Transcript-based media editing | Editing workflow and export options |
| Google Cloud Text-to-Speech | TTS in Google Cloud applications | Voice-specific features and quotas |
Cartesia
We build Sonic, a text-to-speech model for voice applications. You can stream generated audio, select a voice, or use voice cloning with recordings you have permission to use. Try your own text in the playground before integrating the API.
For a complete voice agent, you also need speech recognition and conversation handling. Ink is our speech-to-text model, and Managed Agents is our voice agent builder. Sonic alone does not listen to callers or decide what to say.
Check language coverage, API requirements, and pricing for the model and plan you intend to use. Measure response time in your application: network travel and playback buffering contribute to what a user hears.
OpenAI
OpenAI provides speech and conversational APIs. Compare a speech-to-speech conversation model separately from a text-to-speech endpoint: they give you different control over the words an agent says. Test interruption handling and tool use if you need a complete conversation system.
IBM Watson Text to Speech
IBM Watson Text to Speech is an API for generating speech from text. Include it if you are comparing speech services for an IBM-based deployment. Verify supported languages and pronunciation controls, then test output formats and service limits against your application.
Microsoft Azure Text-to-Speech
Microsoft Azure provides text-to-speech through its speech services. It is a candidate for teams already operating in Azure. Check the selected voice and region, supported Speech Synthesis Markup Language (SSML) controls, and access requirements for custom voices.
Amazon Polly
Amazon Polly is AWS’s text-to-speech service. It is worth evaluating if your application already uses AWS identity and billing. Voice engines differ in language coverage and supported controls, so check the engine you plan to deploy rather than the service-wide feature list.
Speechify
Speechify has reading apps that turn documents and web pages into audio. If your goal is to listen to existing text, start with that workflow. Evaluate its reading, studio, and API products separately; access to one does not tell you what another includes.
Murf AI
Murf AI has a voiceover editor for working from scripts and matching narration to media. Consider it for training materials and recorded presentations. Test how much editing your script needs, and check whether the plan includes the exports and team access you need.
Descript
Descript combines transcription with text-based audio and video editing. Consider it when you want to cut recorded material by editing its transcript. A speech API alone will not replace that editor, so test the recording-to-export workflow before switching.
Lovo AI
Lovo AI is offline. Its parent company, Lovo, Inc., filed for Chapter 7 bankruptcy in May 2026, in the middle of a voice actors’ lawsuit. The Lovo AI alternatives guide covers what Genny users can still recover and which tools to move to.
Google Cloud Text-to-Speech
Google Cloud Text-to-Speech provides speech synthesis through Google Cloud APIs. For an existing Google Cloud application, account and billing integration may simplify adoption. Confirm that your chosen voice supports the language, streaming behavior, and speech controls your application needs.
Test before you switch
Write down which part of Hume you need to replace, and leave time to test before November 13. For speech output, compare delivery on the same text. For an agent, test turn-taking, interruptions, and tool calls. If expression analysis is a requirement, evaluate it separately with a defined task and labeled data; do not infer it from how empathetic a voice sounds.
Compare cost using your expected usage and the features you need. Include retries, regenerated audio, concurrency limits, and commercial rights. A low starting price is not a workload estimate.