AI models with audio input: 25 that listen, not just read

Audio

Services found: 25

#	AI model	API $/1M (in / out)	Context window	Price / mo	Output $/1M	Overview
21	Qwen3.8 Omni Flash Qwen		1M tokens	API only		View
22	MiMo-V2.6-Pro Xiaomi	$0.435 / $0.87	1M tokens	API only	0.87	View
23	MiMo-V2.6-Flash Xiaomi	$0.14 / $0.28	1M tokens	API only	0.28	View
24	MiMo-V2.6-Pro-UltraSpeed Xiaomi	$4.35 / $8.7	1M tokens	API only	8.7	View
25	Perceptron Mk1.5 Perceptron	$0.15 / $1.5	36.9K tokens	API only	1.5	View

Browse by need

Reasoning models169 models

Coding assistants17 models

Vision and images156 models

Long context 1M+82 models

Free tier available80 models

Open weights88 models

Twenty-five models accept audio as input and interpret it rather than merely transcribing it - catching tone, hesitation, overlapping speakers and background sound. Gemini 3.8 Flash, Muse Spark 1.2 and Gemini 3.7 Flash lead, with output from free tiers up to $24 per million tokens.

Audio tokens are billed at their own rate, usually several times the text rate, and duration rather than word count drives the bill. If all you need is a written record, speech-to-text models are cheaper. Use audio understanding where the how matters as much as the what - call quality review, sentiment on support lines, accessibility.

Beyond the wordsTone, hesitation, overlap and background sound, not just the transcript.

Own token rateAudio bills separately and higher than text; duration drives the invoice.

Cheaper alternative existsIf you only need the words, transcription costs a fraction.

Models with this capability

The 25 models below all support this capability. Prices are input and output per 1M tokens from the vendor official price list; for dirhams multiply by 3.6725. Tap a card to open the model page.

Gemini 3.8 Flash

Google DeepMind

in $0.75out $3.751.05M tokens

Muse Spark 1.2

How to choose

Be explicit about what you want inferred — sentiment, intent, speaker state.
Test on noisy real recordings, not studio samples.
Consider transcribing first and reasoning over text when nuance is not required.
Check clip length limits before designing the pipeline.

Common mistakes

Paying audio rates for plain transcriptsThe most common and most expensive mistake in this category.

Over-reading emotionInferred sentiment from audio is indicative, not diagnostic.

Who this suits

Contact centresQuality review and escalation detection.

ResearchInterview analysis where delivery matters.

AccessibilityDescribing audio environments, not just speech.

Related filters

Speech-to-textTranscription with diarisation and timestamps

Realtime voice7 models for live voice conversation

Video understanding53 models that watch and explain video

MultilingualAll 239 models work beyond English

All 239 modelsThe full rating with filters and prices

Frequently asked questions

How is this different from transcription?

Transcription produces text. Audio understanding reasons about how something was said, not just what.

Why is audio billed differently?

Audio tokens have their own rate, several times the text rate, and duration drives the bill rather than word count.

What if I only need the words?

Use a speech-to-text model — dramatically cheaper for the same recording.

Can it identify who is speaking?

It separates speakers; naming them requires your own enrolment step.

Does music or noise break it?

It degrades gracefully, but heavy noise still costs accuracy.

Ratings 1

Avg. score 5

Rate this page

Help us improve