What is Hume AI?
Hume AI, based in New York, is now an audio analysis company. Its Audio Decisions API, launched on 2 November 2026, takes one MP3 or WAV file of 1 second to 5 minutes and up to 25 questions you write, such as "How frustrated does the caller sound?" or "Is there background music?". It returns probabilities, not text: a yes/no, a choice, a score on a scale, or the time spans where an event such as a laugh happens. Each probability estimates the share of human listeners who would give that answer. Hume says the model was trained on tens of millions of listener judgements and agreed with a panel's average 70.7% of the time, against 56.5% for a single human rater.
Hume's voice products are gone. Its Empathic Voice Interface (EVI) for voice agents and its Octave text-to-speech API closed on 13 November 2026 at 12:01 a.m. EST, under a notice posted on 2 October, and Hume says any data left in those accounts is permanently deleted. Saved and cloned Octave voices could not be exported. The developer docs now list only Audio Decisions and the older Expression Measurement API. In January 2026 founder Alan Cowen and several engineers joined Google DeepMind under a licensing deal; Hume stayed independent, led by Andrew Ettinger.
Is Hume AI worth it?
A transcript of a support call tells you what the customer said. It does not tell you that their "I'm fine" came out flat and tired, or that a dog was barking and the agent talked over them. Getting those answers has usually meant stitching together separate models for speaker counting, emotion, language and background noise.
Hume's Audio Decisions API asks all of that in one request. You write the questions in plain English, attach up to 5 minutes of audio, and get back numbers your code can act on: the probability that the caller sounds frustrated, the seconds where someone laughs. It never transcribes, so it pairs with a speech-to-text service such as AssemblyAI rather than replacing one. Hume says it beat Gemini 3.8 Flash, gpt-audio-1.5 and others on 10 benchmarks with an average of 0.82, though those are Hume's own tests.
The price is the attraction. Hume's billing docs put a 5-minute support call with six questions at about 26,400 input tokens, which at $0.47 per million is just over a cent. Watching a live call in 5-second clips comes to about 6,400 tokens a minute. Answers are not billed, and neither are rejected requests.
The catch: this is a v0 API behind a waitlist, from a company in the middle of a reinvention. On 2 October Hume gave developers six weeks before EVI and Octave stopped, with account data deleted afterwards and cloned voices impossible to export. Hume's docs also say the model can be updated without an API change, so answers may shift under thresholds you have tuned. The privacy terms leave gaps for a product that takes in people's voices. Hume's API Data Usage Policy says it does not use customer-submitted data for training its own models, but that page is undated and does not name this product, the general privacy policy (last updated February 2025) excludes API content, and the docs do not say how long uploaded audio is kept. Each request is logged as a run that stores the request and the audio's size; whether the audio itself is stored is not published.
Our take: for scoring a test set of synthetic voices or sampling a call archive, the cost is low enough that the risk is small; as the emotion layer of a live product, it needs a signed data processing agreement and a plan for the day Hume changes direction again.
Where it earns its place
A voice AI team testing text-to-speech models before launch
Score thousands of generated clips for naturalness, glitches and mispronunciation
Send each 10-second clip with three questions; Hume's billing docs estimate about 1,000 tokens a clip, so 10,000 clips is about 10M tokens, or $4.70. Hume pitches this as feedback for training voices too.
A quality lead at a contact centre
Find every call where the customer sounded frustrated and the agent talked over them
Run the same questions over a sample of recorded calls, splitting long calls into 5-minute parts with the SDK, and route high scores to human review. Get a data processing agreement from legal@hume.ai before sending customer voices.
A developer building a voice agent on another provider's speech stack
Give the agent's LLM cues about tone, not just the transcript
Send 5-second clips of the caller alongside speech-to-text and pass the probabilities to the LLM with business rules. Hume's docs say most questions need 5 seconds or more of audio to hold up.
Before you commit
- Setup
- Join the waitlist; once granted, create an API key on the Hume Platform and send a multipart POST to https://api.cloud.hume.ai/v0/decisions/file with the X-Hume-Api-Key header, or install the hume-audio-decisions Python or @humeai/audio-decisions TypeScript SDK.
- Integrations
- REST over HTTP only, Python and TypeScript SDKs, an Audio Decisions skill for Claude Code, Codex and Cursor, and a docs MCP server. No transcript output; pair it with a speech-to-text API.
- Data and privacy
- Hume's API Data Usage Policy says it does not use customer-submitted data for training its own models. Retention of uploaded audio for Audio Decisions is not published; runs record the request and audio size. SOC 2 report on request. EVI and TTS account data was deleted after 13 November 2026. Which AI tools train on your data?
- Cost at scale
- $0.47 per 1M input tokens. Hume's estimates: a 5-minute call with six questions is about 26,400 tokens; live monitoring in 5-second clips about 6,400 tokens a minute. Volume pricing via sales.
- Limits
- One MP3 or WAV of 1 second to 5 minutes and up to 25,000,000 bytes per request; up to 20 choice, score or yes/no questions and 5 span questions; context up to 1,000 characters; 600 requests a minute per organisation.
What’s good
- Plain-language questions about audio, answered as probabilities in one call
- About a cent for a 5-minute call with six questions, at $0.47 per 1M tokens
- Hears tone, overlap and background sound a transcript misses
- 600 requests a minute per organisation; Python and TypeScript SDKs
- Says it does not train on customer-submitted API data
What’s not
- EVI and Octave TTS shut on 13 November 2026, with account data deleted
- Waitlist access and a v0 API whose answers can shift with model updates
- Retention of uploaded audio is not published
- No transcript, so you still need a separate speech-to-text service
- Benchmark claims are Hume's own
Best for
Developers evaluating synthetic voices or sampling call recordings who want tone and sound judgements as numbers, and can live with a waitlisted v0 API.
Hume AI alternatives
Hume AI: common questions
- Is Hume AI worth it?
- Hume's question-and-answer API for audio is clever and cheap, but a company that switched off its voice products with six weeks' notice has to earn back trust before you build on it. We rate it 3.4 out of 5.
- How much does Hume AI cost?
- Audio Decisions: $0.47 per 1M input tokens, no commitment, with a 1M-token trial credit once access is granted. Access is by waitlist. Enterprise volume pricing and VPC, on-premises or in-country deployment on request. EVI and Octave TTS closed on 13 November 2026.
- What are the downsides of Hume AI?
- EVI and Octave TTS shut on 13 November 2026, with account data deleted; waitlist access and a v0 API whose answers can shift with model updates; retention of uploaded audio is not published; no transcript, so you still need a separate speech-to-text service; benchmark claims are Hume's own.
- Who is Hume AI best for?
- Developers evaluating synthetic voices or sampling call recordings who want tone and sound judgements as numbers, and can live with a waitlisted v0 API.
- How much does Hume AI cost for a team or in production?
- $0.47 per 1M input tokens. Hume's estimates: a 5-minute call with six questions is about 26,400 tokens; live monitoring in 5-second clips about 6,400 tokens a minute. Volume pricing via sales.
- Is Hume AI safe for company data?
- Hume's API Data Usage Policy says it does not use customer-submitted data for training its own models. Retention of uploaded audio for Audio Decisions is not published; runs record the request and audio size. SOC 2 report on request. EVI and TTS account data was deleted after 13 November 2026.
- What are the best Hume AI alternatives?
- The closest alternatives we have reviewed are AssemblyAI and ElevenLabs.
Answers checked Nov 16, 2026.