Models
Choose between Ultrafield's proprietary models and open-weight options such as Whisper large-v3, weighing accuracy, latency, languages, and cost.
Ultrafield exposes model choice because speech workloads do not share one universal optimum.
Model families
| Family | Role |
|---|---|
| Ultrafield speech models | Proprietary models and routes optimized by Ultrafield for production inference |
| Whisper large-v3-turbo | Open-weight ASR option suited to workloads prioritizing speed and efficiency |
| Whisper large-v3 | Open-weight ASR option for teams evaluating the full large-v3 accuracy profile |
| pyannote-based processing | Open-weight speaker diarization and related speaker processing |
| GPT-OSS-120B | Hosted language-model processing after transcription |
Exact model identifiers, supported languages, limits, and availability are supplied with API access.
Choose with your own evaluation set
Compare models on:
- Accuracy for your languages, vocabulary, and acoustic conditions
- Batch throughput or streaming latency
- Timestamp and formatting needs
- Speaker overlap and channel characteristics
- End-to-end processing cost
Use the same downstream processors in each evaluation. Changing both the ASR model and the post-processing chain at once makes the result difficult to interpret.
Avoid accidental model drift
Pin an explicit model version for production workflows that require reproducibility. Move to a newer route after evaluating it against the same representative set.
Speech to text
Transcribe batch and streaming audio through one speech-to-text API, with an explicit model choice and optional downstream processing stages.
Streaming
Design low-latency transcription around interim events, finalized segments, backpressure, and reconnection, for live captions and voice agents.