Quickstart

An Ultrafield integration starts with the output your product needs, then adds only the processing stages required to produce it.

1. Describe the workload

Before choosing a model, capture:

• Input type: uploaded files, remote audio, or a live stream
• Languages and expected code-switching
• Typical and maximum audio duration
• Batch throughput or streaming concurrency
• Acoustic conditions such as telephone audio, far-field microphones, or overlapping speech
• Required output and acceptable latency

These details determine the useful accuracy, speed, and cost trade-offs.

2. Choose a speech model

Start with an Ultrafield model or a supported open-weight route such as Whisper large-v3-turbo or Whisper large-v3. Evaluate against representative audio rather than a clean demo clip.

See Models for the factors that guide selection.

3. Add processors

A minimal request returns a transcript. Add processors when the product needs them:

{
  "audio": "<file, URL, or stream>",
  "model": "<speech-model>",
  "processors": [
    "diarization",
    "formatting",
    "entities",
    "sentiment"
  ]
}

This object illustrates composition only; it is not a published request schema.

4. Consume the result

Batch work completes as a job and can notify your application through a webhook. Streaming work emits incremental transcript events and a finalized result. Either path can return stage-specific metadata so your application can distinguish transcript, speaker, and understanding output.

5. Evaluate the whole pipeline

Measure more than word error rate:

• First and final transcript latency
• Speaker-turn accuracy on overlap and short turns
• Formatting and entity precision
• Downstream task quality for LLM processing
• Total processing cost for the actual model mix