LLM processing
Chain speech-derived data into hosted language models for extraction, classification, summaries, and custom transformations.
Language-model processing extends a speech pipeline after transcription and speaker processing. Use it when deterministic processors are not enough for the output your application needs.
Hosted models
Ultrafield hosts models including GPT-OSS-120B for downstream processing. An upcoming gateway will also connect to third-party hosted models from providers such as OpenAI and Anthropic.
Common tasks
- Structured extraction into an application-defined shape
- Conversation or speaker-aware summaries
- Classification and routing
- Topic and intent analysis
- Transformation of transcript segments into downstream records
Keep the chain observable
Record which transcript version, speaker output, prompt or task definition, and model produced a result. This makes evaluation and reprocessing possible when any stage changes.
Control latency and cost
Do not send more context than the task needs. For streaming experiences, keep transcript delivery independent from slower LLM work and update the interface when derived results arrive.
Sentiment analysis
Add conversation-level or segment-level sentiment signals to a speech pipeline, scored from the transcript alongside speakers and entities.
Processing pipelines
Compose speech recognition, speaker processing, audio intelligence, and language models into one observable request instead of stitching vendors.