--- title: Hume launches Audio Decision API to read tone, not just... description: Hume has released the Audio Decision API, which answers plain-English questions about a piece of audio. You can ask things like how many speakers are present,... canonical: https://daily.dev/posts/hume-launches-audio-decision-api-to-read-tone-not-just-transcripts-j8j6ugffi twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Hume launches Audio Decision API to read tone, not just transcripts | daily.dev og:description: Hume has released the Audio Decision API, which answers plain-English questions about a piece of audio. You can ask things like how many speakers are present,... og:url: https://daily.dev/posts/hume-launches-audio-decision-api-to-read-tone-not-just-transcripts-j8j6ugffi og:image: https://api.daily.dev/og/posts/J8J6Ugffi.png og:image:alt: Hume launches Audio Decision API to read tone, not just transcripts og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Hume launches Audio Decision API to read tone, not just transcripts **[Collections](https://daily.dev/sources/collections)** · 2 min read · 0 upvotes · 0 comments ## Summary Hume has released the Audio Decision API, which answers plain-English questions about a piece of audio. You can ask things like how many speakers are present, whether anyone is laughing, or if a caller sounds frustrated, and the API returns a probability for each possible answer. It works directly from audio without needing a transcript. That matters because transcript-based voice agents miss cues like sighs and pauses, for example treating a flat 'I'm fine' at face value when the tone says otherwise. ## Content Hume AI has launched the Audio Decision API, an audio-native model that answers questions about a recording by listening to the audio itself. It does not work from a transcript. ## How it works You send audio and write your own questions and answer options in plain English, such as "Is the caller frustrated?" The API returns a probability for each answer you defined. Questions can also cover things like how many speakers there are or whether anyone is laughing. The model takes sighs, laughter, hesitation, overlapping speech and background sounds as input. A transcript flattens all of that into the same words. A sarcastic "Oh, perfect" reads as happy in text, and a flat "I'm fine" reads as fine. A model that hears the delivery can tell the difference. ## Accuracy Hume says the model was trained on more than 10 million human judgments. In initial tests, it agreed with the average answer of a listener panel 70.7% of the time. A single listener agreed with that panel average 56.5% of the time. The API is estimating how human listeners would answer, so that's the benchmark that matters. ## Why it matters The obvious use is business: call centers and support lines that need to know whether a caller is upset while the call is still going. @omarsar0 thinks the bigger opening is personal agents, since more of us talk to them by voice. An agent could ask "is this person frustrated?" in one call and change how it responds to what you said and how you said it. @kloss_xyz put it more bluntly: labs have spent years making AI voices sound human, while the agents on the other end were reading transcripts and never really listening. I think that's a fair description of the gap. Most voice agents today see the word "fine" and move on. The 70.7% figure is a start, not a guarantee, and it's an average over human opinion, which is itself noisy. But a probability per answer is something an app can act on, and that's more useful than a transcript. ## Questions this post answers ### How does Hume's Audio Decision API work without a transcript? You send it an audio clip plus a plain-English question, such as whether a caller sounds frustrated or whether anyone is laughing, and it returns a probability for each possible answer. It analyzes the audio directly, picking up cues like sighs and pauses that a transcript would miss, such as a flat 'I'm fine' that actually signals frustration. _Building a voice agent that needs to catch tone, not just words? daily.dev surfaces releases like this as they land._ ### Why do transcript-based voice agents miss signs of caller frustration? Because they only process the text of what was said, so a phrase like 'I'm fine' reads as neutral even when a sigh or pause before it signals the opposite. Tone carries much of what a frustrated caller communicates, and a system that only reads transcripts misses those cues entirely. _Developers debugging tone-blind voice agents can follow tools addressing this gap on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 2 discussions and 26 comments across x (as of 2026-11-03). **TL;DR:** Reaction centers on the appeal of judging tone straight from audio instead of a lossy transcript, with repeated curiosity about the 70.7% vs 56.5% accuracy numbers, mixed with jokes about surveillance-style use in call centers and skepticism about noisy or crosstalk-heavy recordings. **Sentiment:** 35% positive · 45% mixed · 20% skeptical **The case for** - Detecting sighs, hesitation, and sarcasm directly from audio solves something transcripts genuinely miss. - Beating a single human rater's agreement rate (70.7% vs 56.5%) is seen as a meaningful, concrete improvement for call QA. - A single audio-native call that scores every recording (vs. manual spot-checks) is viewed as a practical efficiency gain. **The pushback** - Concern that companies could use frustration detection to manage customers' emotions rather than actually help them (e.g., still put them on hold). - Skepticism about how well it will handle noisy lines, accents, and overlapping/crosstalk speech. - Accuracy figures are Hume's own self-reported eval, not independently verified. **By community** - x (mixed): People are intrigued by audio-native tone detection and the accuracy stats, but temper that with jokes about customer-service misuse and questions about real-world robustness. **Hottest debate:** Whether this is a genuinely useful QA tool or mainly enables companies to detect customer frustration without actually resolving it. **Open questions** - How does it perform on noisy calls, bad phone lines, or background machinery noise? - How does it handle overlapping speech when multiple people talk at once? - How much would results vary between different individual human raters on the same clip? - Can it detect whether a voice is synthetic/real rather than just emotional tone? **Highlights** > @testingcatalog Now customer service can hear you’re frustrated and still put you on hold > — [iambokensha on x · 1 comments](https://x.com/iambokensha/status/2117342709572620430) > @testingcatalog as a Java backend guy i like one call over 10 models, same recording should give the same answer when... > — [dusangran on x](https://x.com/dusangran/status/2117314882794348697) > @testingcatalog This matters for call QA. Today a team spot-checks a few recordings by hand. A model that agrees with the panel more often than any one rater can score every call, and Hume's demo ran 25,642 files at about 180 ms each. > — [DoDataThings on x](https://x.com/DoDataThings/status/2117338793762230319) > @testingcatalog a frustrated caller is exactly who a small repair shop has to call back first. would like to see how it handles a bad line from someone standing next to a running machine > — [FlorianAure on x](https://x.com/FlorianAure/status/2117342086290940052) > @testingcatalog 70.7% agreement with a human panel is solid. does "is this voice real?" count as a question it can answer, or is that out of scope? > — [Frelzu on x](https://x.com/Frelzu/status/2117313882339311836) **Source threads** - [x](https://x.com/svpino/status/2117474547373130116) · 0 points · 0 comments - [x](https://x.com/testingcatalog/status/2117312236993544277) · 1 points · 26 comments ## Similar posts on daily.dev - [Your AI Voice Agent Is a Black Box. Here’s How to Open It.](https://daily.dev/posts/your-ai-voice-agent-is-a-black-box-here-s-how-to-open-it--0slfkjjqy) · ITNEXT · 0 upvotes · 0 comments - [Your AI Answer Was a Guess. Here’s Why.](https://daily.dev/posts/your-ai-answer-was-a-guess-here-s-why--soykln9ad) · Medium · 1 upvotes · 0 comments - [AI-Enabled Advisory Services for Higher Education](https://daily.dev/posts/ai-enabled-advisory-services-for-higher-education-wps7f1zel) · databricks · 0 upvotes · 0 comments --- Tags: [#voice-ai](https://daily.dev/tags/voice-ai) [View this post on daily.dev](https://daily.dev/posts/hume-launches-audio-decision-api-to-read-tone-not-just-transcripts-j8j6ugffi) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Hume launches Audio Decision API to read tone, not just transcripts","url":"https://daily.dev/posts/hume-launches-audio-decision-api-to-read-tone-not-just-transcripts-j8j6ugffi","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/hume-launches-audio-decision-api-to-read-tone-not-just-transcripts-j8j6ugffi"},"datePublished":"2026-11-02T16:32:39.069Z","dateModified":"2026-11-03T11:32:08.888Z","description":"Hume has released the Audio Decision API, which answers plain-English questions about a piece of audio. You can ask things like how many speakers are present,...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/hume-launches-audio-decision-api-to-read-tone-not-just-transcripts-j8j6ugffi","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"voice-ai","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Hume launches Audio Decision API to read tone, not just transcripts"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/hume-launches-audio-decision-api-to-read-tone-not-just-transcripts-j8j6ugffi#faq","mainEntity":[{"@type":"Question","name":"How does Hume's Audio Decision API work without a transcript?","acceptedAnswer":{"@type":"Answer","text":"You send it an audio clip plus a plain-English question, such as whether a caller sounds frustrated or whether anyone is laughing, and it returns a probability for each possible answer. It analyzes the audio directly, picking up cues like sighs and pauses that a transcript would miss, such as a flat 'I'm fine' that actually signals frustration. Building a voice agent that needs to catch tone, not just words? daily.dev surfaces releases like this as they land."}},{"@type":"Question","name":"Why do transcript-based voice agents miss signs of caller frustration?","acceptedAnswer":{"@type":"Answer","text":"Because they only process the text of what was said, so a phrase like 'I'm fine' reads as neutral even when a sigh or pause before it signals the opposite. Tone carries much of what a frustrated caller communicates, and a system that only reads transcripts misses those cues entirely. Developers debugging tone-blind voice agents can follow tools addressing this gap on daily.dev."}}]} ```