Hume Audio Decision API Review: $0.47 Per Million Tokens

Contact centers score every call against a rubric, and the rubric runs on the transcript. The transcript has the words. It does not have the pause before the answer, the sigh, or two people talking over each other, which is often where a call actually goes wrong. Hume's Audio Decision API, launched on November 2, 2026, takes a different route: send it the raw recording and a question in plain language, and it answers with a probability instead of text.

No transcript comes back, and no explanation either. What you get is a number your code can act on, such as an 87 percent chance the caller sounds frustrated, or the time spans where someone laughed. It does less than a transcription service on the surface, and something transcription cannot do underneath: it judges how audio sounds, not only what it says.

What the Audio Decision API actually does

Hume calls it the first audio-native model that answers questions about speech and sound. That "first" is the company's own description, and the architecture is different from the usual pipeline in a specific way. A normal voice stack turns speech into text, then hands the text to a language model. Here the raw audio goes in and a structured answer comes out, with no transcript in between. The model was trained on more than 10 million human judgments about how people perceive speech and sound, according to Hume, and its answers estimate how a panel of human listeners would judge the same recording.

The docs list the jobs it fits: evaluating text-to-speech output, testing voice agents, reviewing support calls, and labeling speech corpora. Hume's launch post adds two more: searching an archive of recordings for the ones that match a question, and tracking how expression changes across a call. The company says the API is fast enough to ask dozens of questions every few seconds on a live call, and to make an archive of 40,000 recordings searchable in minutes. Both figures are Hume's claims, with no independent test published yet.

It is not a voice agent, not a text-to-speech service, and not a transcription tool. It produces judgments, not words. Voice models that generate speech, like Block's Dessa, a voice AI that learned from listening, work the other side of the same problem. If your workflow needs a transcript, Hume's own diagram pairs the Decision API with a speech-to-text service and a language model, so the transcript and the audio judgments arrive together and the language model weighs both before acting.

Four question types, one request

Every question has a type, called a primitive in the docs, and one request can mix them. The four types cover most of what a team would ask about a recording.

Hume's docs walk through a simulated support call with two questions: whether the agent talks over the caller, and how natural the agent sounds on a five-point scale. The answers came back as 0.1274 for talking over, meaning about 13 percent of raters would say yes, and a naturalness score of 2.72 out of 5, with the full probability spread across the five levels.

Three rules from the docs matter for anyone building on it. Ask each question for one focused judgment, then combine the answers in your own code with thresholds you choose. Send all the questions about a recording in one request, and keep the same questions in the same order for every recording you compare, because an answer can shift when the other questions change. And add context to the request, such as who is speaking and what the call is about, because the docs say it can improve performance.

What it costs

The published price is $0.47 per million input tokens, and answers are not billed. Approved users get 1 million free input tokens to start. That is the only published price for this API: there is no monthly plan attached to it, and Hume's older subscription tiers belonged to the text-to-speech and voice products that are shutting down, not to this one.

In pounds and Canadian dollars, $0.47 per million tokens works out to about £0.35 and C$0.66 at November 10, 2026 market rates (1 USD = 0.7436 GBP, 1.4117 CAD). Those are approximations, and the dollar figure is the one Hume bills.

What a token buys depends mostly on the audio itself. The length of the recording makes up most of the tokens, so ten questions about one file cost little more than one question. Hume's own docs example, a short simulated call with two questions, used 11,234 input tokens. At the list price that is about half a cent per call, and the 1 million free tokens cover roughly 89 calls of that size. That is our arithmetic on Hume's example, not a published rate, and longer calls will use more.

Before running a batch, the API has an endpoint that estimates the input tokens for a file, so you can price a job before you pay for it. One more piece of fine print from the docs: the model can be updated without a change to the API version, which means answers for the same audio and the same questions can change after an update. And the run endpoints return a run's status and request, but not its answers, so store the run ID and the answers you need yourself.

How to try it

Access is by request. Join the waitlist on Hume's site, and while you wait, the playground lets you try the real-time and file-upload experience with your own questions and no setup.

Once you have a key, a request is a multipart POST to https://api.cloud.hume.ai/v0/decisions/file, with the key in an X-Hume-Api-Key header. The request has two parts: a JSON part with your questions and optional context, and the audio file itself. One file per request, MP3 or WAV, from one second to five minutes long, heard as 16 kHz mono audio. Python and TypeScript SDKs cover every endpoint, read the key from a HUME_API_KEY environment variable, retry on rate limits, and split recordings longer than five minutes into one request per part. For live calls, send clips of at least five seconds as they arrive.

Hume's example request asks two questions about a simulated test call:

curl https://api.cloud.hume.ai/v0/decisions/file \
  -H "X-Hume-Api-Key: $HUME_API_KEY" \
  -F 'request={
    "context": "A simulated test call in which a voice AI agent helps a caller reschedule an appointment. Both voices are synthetic, and the agent speaks first.",
    "questions": {
      "talks_over_caller": {
        "type": "noul",
        "instructions": "Does the agent talk over the caller?"
      },
      "agent_naturalness": {
        "type": "score",
        "instructions": "How natural does the agent sound?",
        "criteria": ["Not at all", "2", "3", "4", "Extremely"]
      }
    }
  };type=application/json' \
  -F file=@test_call.wav

The response keys each answer by the ID of its question, carries the run ID of the model run, and reports the input tokens the request is billed, under "received".

What the benchmarks do and do not prove

Hume reports that its model agreed with a human panel's average judgment 70.7 percent of the time, against 56.5 percent for a single human rater. It also reports an average score of 0.82 across 10 benchmarks, with a median latency of 0.23 seconds, and a chart that puts it ahead of Gemini 3.8 Flash, Qwen3.8-Omni-Flash, Inkling-Small, gpt-audio-1.5, Gemini 3.5 Flash Lite, and Nemotron 3 Nano Omni. Every one of those figures comes from Hume. No independent replication exists yet, and Hume has not said how many recordings or raters were used, or in which languages.

The honest way to read the 70.7 percent figure is that the model matches a panel average better than one person does, which is a low bar with a specific meaning. Panels average out individual quirks, so beating a single rater says the model learned the consensus, not that it is right in any absolute sense. And a probability near 0.5 does not mean the model is unsure. The docs say it means human raters would split on that audio, which is a statement about the recording, not the model.

Hume itself draws a line between objective and perceptual questions. Counting speakers or detecting a siren is relatively objective. Whether a voice sounds frustrated, empathetic, or natural is a matter of perception, and listeners do not always agree. The launch post also tells teams to validate their evaluations against representative human ratings, which is the right advice and worth following literally: run 50 of your own calls through it before you trust any threshold with a real decision.

Hume vs Tazo vs Smaug-Decisions

Hume is not the only team building this category. Three audio decision models arrived within days of each other in early November, and they aim at different jobs. Each company tested its own model on its own benchmark, so the numbers below cannot be compared with each other, a warning The Tool Nerd states plainly in its November 10 roundup.

The prices are in different units, tokens against hours, so no honest "cheapest" claim follows from the table alone. The real split is about where the decision happens: Hume and Tazo run in the cloud, while Smaug runs on your own hardware or phone, inside an app Abacus.AI calls Cascade.

The bigger Hume story: the voice products end November 13

The launch lands in the same month Hume shuts down the products it was known for. On October 2, Hume said its text-to-speech and Empathic Voice Interface APIs would end access on November 13, 2026 at 12:01 a.m. EST, with account data permanently deleted afterward. By November 9, the old text-to-speech docs and changelog addresses redirected to the Audio Decisions docs, the homepage no longer mentioned the Octave voice model, and the pricing page redirected to the homepage, according to Versely Studio's check that day.

The Decision API is a separate product line, and Hume has not described a direct migration path from the old services. If you built on EVI or Octave, this API does not replace what you had: it judges audio, it does not generate speech. Voice-agent builders who need a spoken voice will need a new provider, and Mistral's Voxtral 3, a phone-ready voice agent at $0.04 a minute, is one of the current options.

Hume is a New York emotion-AI research company founded in 2021 by Alan Cowen, a former Google DeepMind scientist, which raised a $50 million Series B in 2024, according to APIRank's review. The pivot is a bet that judgment is worth more than generation: a model that tells you what happened in a call is harder to replace than one more voice.

Who should use it, and who should wait

Use it if you run a contact center and score calls against a rubric, if you test voice agents and need numbers for naturalness, interruptions, and tone instead of impressions, if you sit on an archive of recordings nobody has time to listen to, or if you label speech data and want a first pass before humans check it. For the call-analysis side, our roundup of the best AI meeting note-takers of 2026 covers the tools that already summarize calls, and Hume fits beside them as the layer that scores how a call sounded.

Wait if you need access today, because the waitlist is the only door in. Wait if you need decisions on clips shorter than five seconds in a live call, or independent proof of accuracy before a workflow depends on it, or decisions on the device, where the free Smaug models are the honest option. And if you need a transcript as well, Hume's own workflow pairs the Decision API with transcription rather than replacing it.

Pro tips and common mistakes

• Ask one focused judgment per question, then combine the answers in your code with thresholds you choose. A question like "did the agent handle the call well" depends on several things, so ask about each one separately.
• Keep the same questions in the same order for every recording you compare, because an answer can shift when the other questions change.
• Treat a probability near 0.5 as "human raters would split on this audio", not as the model being unsure.
• Store the answers yourself. The run endpoints return a run's status and request, but not its answers.
• Estimate cost first with the input_tokens endpoint, and remember that ten questions about one file cost little more than one, because the audio dominates the token count.
• Do not read a perception score as a fact about what a speaker feels. The model estimates how listeners would judge the audio, not a person's inner state.
• Re-check your thresholds after Hume updates the model, because the model can change without a version bump.

How much does Hume's Audio Decision API cost?

$0.47 per million input tokens, and answers are not billed. Approved users get 1 million free input tokens. There is no monthly plan for this API. In pounds and Canadian dollars that is about £0.35 and C$0.66 per million tokens at November 10, 2026 market rates.

Is the Audio Decision API available now?

It launched on November 2, 2026, but access is by request through a waitlist. A playground with real-time and file-upload testing is open, and you can add your own questions there with no setup. Hume has not said when general access opens.

What audio files does it accept?

One MP3 or WAV file per request, from one second to five minutes long, heard as 16 kHz mono audio. The Python and TypeScript SDKs split longer recordings into one request per part. For live calls, send clips of at least five seconds as they arrive.

How accurate is it?

Hume reports 70.7 percent agreement with a human panel's average judgment, against 56.5 percent for a single human rater. Those figures come from Hume's own evaluation, with no independent replication published yet and no disclosure of how many recordings or raters were used.

Does it replace transcription?

No. It returns probabilities, not words, so it cannot tell you what was said. Hume's own workflow pairs it with a speech-to-text service and a language model, so the transcript and the audio judgments reach the language model together before it acts.

What happens to Hume's TTS and EVI APIs?

Access ends on November 13, 2026 at 12:01 a.m. EST, and Hume says account data is permanently deleted afterward. The Audio Decision API is a separate product, and Hume has not described a migration path from the old services to the new one.

The Audio Decision API stands out among this month's voice launches because it does not talk. It turns the parts of a call a transcript throws away into numbers a program can use, at a price that makes bulk analysis cheap. The open question is accuracy, and only your own recordings can answer that. Join the waitlist, run 50 of your own calls through the playground, and set your thresholds from what you see, not from Hume's chart.

Sources

• Hume, "Introducing Hume's Audio Decision API: A decision layer for speech and sound" (November 2, 2026)
• Hume API docs, "Audio Decisions API" (updated November 7, 2026)
• Hume, "Decision Model API" product page
• DecisionModels.ai, "Hume Audio Decisions API" (price checked against Hume billing docs, November 3, 2026)
• FastToolHub, "Hume Audio Decision API"
• The Tool Nerd, "13 AI Tools I Discovered This Week" (November 10, 2026)
• AlphaSignal, "Hume Opens Its Audio Decision API to Answer Any Question About Sound" (November 2, 2026)
• Synthship, "Hume's Audio Decision API answers your own questions about a recording" (November 4, 2026)
• Superpower Daily, "Hume introduces an API for asking questions about sound, with access by waitlist" (November 2, 2026)
• Versely Studio, "Hume Octave 2: API Ends 13 Nov 2026, What to Use"
• APIRank, "Hume AI 2026: The Empathic Voice and OCTAVE TTS Review for Production Voice Agents"
• HCI Insights for November 3, 2026