Kyma now hears
If your agent has been blind to audio, that just changed. Two new endpoints went live on Kyma today, and together they let any AI agent listen to a clip the way a person would: not just hear the words, but read the room.
What
/v1/audio/transcriptions returns the spoken text. /v1/audio/understand answers free-form questions about everything else in the audio - tone, music, ambient sound, language, speaker emotion. Both go through the same API key you already use for chat, image, and video. Both bill per minute, rounded up.
The first endpoint is OpenAI Whisper API compatible, so any tool that already knows how to talk to OpenAI's transcription endpoint works against Kyma with one base-URL change. The second has no equivalent on other gateways - it's the part of audio understanding that a transcript drops on the floor.
Why this matters
There has always been a heavy way to do audio understanding: feed the file to a multimodal LLM and pay for every minute as if it were context tokens. It works. It's also about fifty times more expensive than it needs to be.
A typical 1-hour clip on a multimodal frontier model runs around $5 and 30-60 seconds of latency. The same hour through Kyma's two new audio endpoints costs under ten cents, in about ten seconds:
| Approach | Cost / 1-hour clip | Time |
|---|---|---|
| Multimodal LLM on the full audio | ~$5 | 30-60s |
| Kyma audio endpoints | < $0.10 | ~10-15s |
The trick is that audio is not one problem. What was said is a transcription problem. What it sounded like is a scene-understanding problem. Each piece has a much faster, much cheaper specialist tool. Kyma now serves both, with a single key.
For an agent watching dozens of videos in a session - research, monitoring, content moderation, archival search - the savings compound from "noticeable" to "this changes which workflows are economic at all."
Where this fits - giving your agent eyes and ears for any social video
The launch ships with a companion CLI we open-sourced today: watch-cli.
watch-cli is one command that takes a URL - YouTube, X, LinkedIn, TikTok, Reddit, Vimeo, Facebook - and gives an agent everything it needs to "watch" the video without ever calling a vision model on the full file. Under the hood it's yt-dlp for the download, ffmpeg for evenly-spaced frames and the audio track, and Kyma's two new endpoints for the transcript and the audio scene. The output is structured for an agent to consume in one shot: a video path, a list of frame JPGs, the transcript, and a one-paragraph audio scene description.
Drop it into Claude Code or any agent harness with a single line in the system prompt - "you have a watch command that takes a URL" - and the agent suddenly has video understanding. We've been using it internally to triage launch announcements, monitor competitor videos, and pull notes out of recorded talks. The cost so low that nobody thinks twice about pointing it at a 30-minute keynote.
This is the workflow we built the audio endpoints for. If you want the install steps and the full command surface, the watch-cli README has it.
How to use it
The minimum viable example - a single sync call, no polling, no async dance:
curl -X POST https://kymaapi.com/v1/audio/transcriptions \
-H "Authorization: Bearer $KYMA_API_KEY" \
-F "file=@meeting.mp3" \
-F "model=transcribe"
model: "transcribe" is an alias - it always points to the current best ASR on Kyma. When a faster Whisper or a stronger ASR lands, your code doesn't change. Same idea for model: "audio-understand" on the scene-understanding endpoint.
For multipart parameters, the OpenAI SDK shape, and the audio understand endpoint, the docs are concrete: audio family overview, transcribe reference, understand reference.
What's next
- •Bigger audio inputs. The 25 MB inline cap covers ~30 minutes of mono 16kHz mp3, which is enough for podcasts and most calls but not for long-form video. A Files-API path is in design for chunked uploads.
- •Text-to-speech on the same single-key principle, with the same alias pattern -
ttswill resolve to whichever model is current the day you read it. - •Faster ASR. When the next-generation model lands, the
transcribealias rolls forward automatically. Code pinned towhisper-v3-turbokeeps working unchanged.
If you want to try it, grab a Kyma key - the $0.50 of free credit on signup is enough to transcribe and understand several hours of audio before you spend a cent of your own.