Microsoft AI: MAI-Transcribe 1.5

MAI-Transcribe 1.5 is a speech-to-text model from Microsoft AI, released on June 2, 2026. It sits in the transcription category, so it is built for one job: turning spoken audio into written text rather than generating conversation or code. If you are assembling a pipeline that starts with voice, this is the component Microsoft provides for that first step.

The model takes audio as its input and returns a transcription as its output. Microsoft describes it as multilingual, with reliable transcription across 43 languages and diverse audio conditions. The provider notes it is suited for captions, call transcription, subtitling, accessibility, and other voice-enabled applications. The context length is not published in the model metadata, and the maximum output field is listed as zero, so the practical output size is governed by the length of the audio you submit and the service's own handling rather than by a documented token budget.

Those capabilities map onto a fairly clear set of use cases. If you need subtitles or closed captions for recorded or live material, the model's subtitle and caption focus applies directly. Call transcription covers contact centre archives, meeting notes and compliance recordings. Accessibility work, such as making recorded talks or internal videos readable, is another fit. Because the model handles 43 languages, teams serving multilingual audiences can route audio in any of those languages through a single model instead of maintaining a separate recogniser per locale. Voice-enabled applications that need a text layer before downstream processing, such as search over recordings or summarisation, also start here.

You call it through api14's OpenAI-compatible API using the model id "microsoft/mai-transcribe-1.5". Because the endpoint follows the OpenAI format, existing client libraries and request shapes you already use for other models will work without special handling. Note that this model lists no supported request parameters, so you should send the audio with the standard request structure and avoid assuming that sampling controls, temperature or similar knobs will have any effect on a transcription task.

There is not much in the supplied details about hard limits, so treat audio length and throughput as things to verify against the live service rather than assume from documentation. Also consider whether a speech-to-text model is what you need at all: if your task is translation between written languages, or text generation over the transcript, this model only covers the recognition step and you will need another model for the rest of the chain.