Overview
Orpheus is Hecttor's speech-enhancement SDK for machines. Use it anywhere a model is on the receiving end: ASR pipelines, voice agents, real-time transcription, voice-AI products built on Pipecat or LiveKit Agents.
Internally, Orpheus is exposed as ASRSpeechEnhancer across all SDK language bindings.
Optimized for ASR, not for hearing
Orpheus is tuned to improve transcription accuracy of downstream ASR, measured by Word Error Rate (WER) or Character Error Rate (CER). It is not tuned to make audio sound better to humans.
Subjective listening is not a valid evaluation method for Orpheus. Enhanced audio may sound unchanged (or even subtly different) while still significantly lowering WER on your downstream ASR.
Always evaluate against the ASR pipeline you intend to ship. Compare transcription results on identical audio before and after enhancement, and compute WER (or CER) against ground-truth references.
For a step-by-step evaluation protocol, see Evaluations.
When to use Orpheus
Use Orpheus when audio reaches your ASR or voice-AI stack in conditions that hurt transcription:
- Background noise (offices, cafés, streets, calls).
- Low-bitrate or narrowband sources (telephony, low-quality codecs).
- Accented or fast speech where the ASR struggles.
- Real-time conversational agents where every transcription error compounds downstream.
Orpheus runs in real time, in-process, on the same chunk boundaries your ASR pipeline already uses.
Models
Orpheus ships multiple noise-cancellation models across a few architectures, suited to different audio environments. We recommend testing more than one against your dataset; the best choice depends on your audio mix and your ASR.
Available model names and their default blend-weight values are provided during onboarding.
A blend parameter is available to mix original and enhanced audio (0.0 = fully original, 1.0 = fully enhanced). Each model ships with a default tuned for ASR performance. It's a solid starting point, but you can try other values against your evaluation to see if a different blend lifts WER further on your audio.
Where to go next
- Getting Started: install Orpheus and run your first enhancement.
- Examples: common configurations and integration shapes.
- Voice Activity Detection: coming soon.
- Turn Taking: coming soon.