NVIDIA Nemotron 3.5 Transcribe
Transcribe and analyze recorded audio with a downloadable model controlled through natural-language instructions.
Hume Audio Decisions API
www.hume.ai
An audio analysis API that answers custom questions about recordings with structured probabilities and event time spans.
Visit website
Pricing
Paid
Available on
API, Python SDK, TypeScript SDK, Web demonstration
Category
Ai Audio
What is Hume Audio Decisions API?
Hume Audio Decisions API accepts a recording and questions written in plain language, then returns structured probabilities. Questions can concern speech delivery, spoken content or other audible events. It is designed for tasks such as evaluating synthetic speech, testing voice agents, reviewing support calls and labeling speech datasets.
The API returns numerical judgments rather than a transcript, conversational response or explanation. Application code decides how to threshold, rank or combine the answers. Hume's product page offers an interactive demonstration with sample audio, microphone input and uploads, alongside a waitlist for API access.
Key features
Custom audio questions
Ask about how speech sounds, what a speaker says or audible events such as background noise.
Yes/no and choice answers
Receive a probability for a yes/no judgment or a probability distribution across unordered answer options.
Ordered scoring
Define ordered levels for a quality such as naturalness and receive a probability for each level.
Event time spans
Ask when an event occurs and receive intervals where it may be present in the recording.
Typed SDK integration
Python and TypeScript SDKs provide typed requests and responses, retries and helpers for splitting longer recordings.
Input-token estimation
Estimate input tokens before submitting a decision request. Estimation is not billed but counts toward request limits.
Who is it for?
Voice developer evaluating synthetic speech
Compare recordings against a consistent set of naturalness and delivery questions.
Support quality analyst selecting calls
Use custom judgments to identify recordings for closer human review.
Speech researcher labeling recordings
Apply defined questions across a speech dataset and inspect ambiguous results.
Voice-agent tester locating events
Find relevant audible events within test recordings for follow-up analysis.
How to use Hume Audio Decisions API
- Explore the product demonstration and request API access through Hume's waitlist.
- Once access is granted, create an API key in Hume Platform and keep it on a controlled server.
- Install a documented SDK; the Python quickstart requires Python 3.10 or later.
- Prepare an MP3 or WAV recording and define questions using the appropriate answer types. Keep the questions and their order consistent across recordings you compare.
- Send the recording and questions, then interpret the returned probabilities or time spans in your application.
Hume Audio Decisions API pricing
Paid
Usage is billed by input tokens covering audio, questions and context; answers are not billed. Approved organizations receive trial credit. No-commitment usage and Enterprise arrangements are available. Audio duration, including silence, contributes to usage.
Check current pricing ↗What to know before you use it
- API access is offered through a waitlist and access request process; the public demonstration does not establish unrestricted API access.
- Each request accepts one MP3 or WAV recording lasting 5 seconds to 5 minutes, with a combined request-size limit of 25,000,000 bytes. Longer recordings require splitting.
- Requests allow up to 20 yes/no, choice and score questions combined, plus 5 span questions. All API keys in an organization share a 60-request-per-minute limit.
- Probabilities estimate human judgments rather than objective ground truth or model confidence. Answers can change when other questions in the request change, so comparisons should use the same questions in the same order. Model updates can also change answers without an API-version change.
Frequently asked questions
Does it produce transcripts or written explanations?
No. It returns structured judgments for application code to interpret.
Can I analyze recordings longer than five minutes?
Split them into recordings of five minutes or less, keeping each part at least five seconds long. Each part is sent separately.
Can a browser call the API directly?
Hume recommends relaying browser requests through a controlled server so the API key is not exposed in client-side code.
Are estimates and retrieved runs billed?
Input-token estimation, listing runs and retrieving runs are not billed. Successful decision requests are billed for their input tokens.