Changelog

Scaling improvements

Helios, Apollo, and Lyra now have auto-scaling enabled with higher baseline capacity, ensuring consistent performance and lower latency during peak usage. Lyra also adds support for high-throughput scaling for concurrent sessions.

Lyra — policy name in results

Policy result events now include the policy name, making it easier to identify which policy produced each result when using multiple policies in a session.

Lyra — psyche biomarker option

Psyche is now available as an explicit biomarker option in Lyra, allowing you to request psyche-specific analysis during voice agent sessions.

Helios pipeline v1.0.0

Released the first stable version of the Helios pipeline with improved audio processing using streaming silence removal for cleaner input before model inference.

Lyra — demographics endpoint integration

Lyra now integrates with the demographics endpoint, allowing summary field extraction to be configured per policy. This enables dynamic field extraction based on the active policy during a voice agent session.

Lyra — optional user attributes

User attributes are now optional when initiating a Lyra session. Additionally, fixed an issue where summary fields were not being cleared correctly between sessions.

Lyra — speech progress events

Lyra now emits real-time speech progress events during voice agent sessions, providing streaming insight into the conversation as it progresses.

Lyra — custom organisation policies

Voice Agent integrations now support custom policy configurations per organisation, allowing you to define tailored extraction fields, safety analysis triggers, and agent evaluation settings for your specific use case.

Apollo file upload fix

Fixed an issue where Apollo model runs could fail if the two required audio files were not uploaded simultaneously. The API now correctly waits for all files before starting the model run, regardless of upload timing.

Account endpoint — enabled features

The /v1/account endpoint now returns a list of enabled features for your API key, allowing you to programmatically check which models (Helios, Apollo) are available for your account.

Shorter recording support

Helios now accepts shorter voice recordings, with updated audio segmentation for improved handling of brief audio samples.

Japanese language support

Helios and Apollo now support Japanese language audio recordings via the ja-JP language code.

Pipeline speed improvements

Reduced processing times for Helios model runs through optimised pipeline execution and increased memory allocation.

Helios performance improvements

Upgraded the Helios audio embedding model with improved prediction times and reserved processing capacity for more consistent response latency.

Apollo calibration update

Updated Apollo to v1.1 with improved calibration for more accurate disorder probability and severity scores. The model version returned in results is now "apollo_1.1".

Apollo v1.0 released

Apollo is now available for clinical voice-based assessment of depression and anxiety. Apollo requires two audio recordings — a mood question response and a read-aloud passage — and returns:

• Disorder-level results for depression and anxiety, including probability and severity scores
• Granular symptom scores across all DSM-V criteria with severity distributions (none, mild, moderate, severe)
• Clinical report accessible via a secure web view and downloadable PDF, with time-limited access tokens

Audio quality warnings

Model run results now include a warnings field. When audio quality is below the confidence threshold, the warning LOW_AUDIO_QUALITY is included in the response. This helps you identify recordings where results may be less reliable due to poor audio input.

The warnings field is null when no issues are detected.

Helios model versioning

Model run results now include a model field indicating which version of Helios was used. Completed runs return "helios-1.0" in the response. Previously completed runs will show a blank value for this field.