Custom speech

In the Ultra Transcenders books

AI-103

Each book explains Custom speech in context, with comparison tables and the common traps.

Terms in this definition

• Domain

  A way of grouping workspaces by area of the business, in support of a data mesh approach. Items take on their workspace's domain, letting you filter the OneLake catalog by it, and certain tenant settings can be passed to domain admins; domains have no effect on access permissions.

• Fine-tuning

  Continuing to train a pretrained base model on task examples, through supervised fine-tuning, DPO or RFT, so its weights shift towards a style, format or task. It neither acts as a safety control nor adds new knowledge well.

• Foundry project

  A container beneath a Foundry resource that keeps one team's agents, data, files, evaluations and project connections separate. Model deployments and resource-level connections are shared from the parent, and governance is not configured at this level.

Related terms

• Audio + human-labeled transcript data

  Training material for custom speech: recordings in mono 16-bit PCM WAV (RIFF) format at 8 or 16 kHz, zipped up together with a plain-text file listing each recording's name alongside what was said.

• Custom speech endpoint

  Needed for real-time recognition with a custom speech model and identified by its endpointId. Batch transcription doesn't require one, and the model behind it can be replaced without any downtime.

• Custom speech model expiration

  What happens once a custom speech model expires: real-time endpoint calls drop back to the newest base model for that locale, whereas batch jobs naming the model fail with a 4xx error; the model itself remains.

• Speech Studio

  A web portal offering no-code Azure Speech tools, including batch speech to text, audio content creation, custom voice and custom speech.