Prism EQ Pipeline
Research-grade infrastructure for your voice data.
Turn raw recordings into rich data for research and training. Prism EQ is a configurable pipeline for every stage of model development, all in your own cloud.
Pre-training · Mid-training · Post-training · Fine-tuning
Find the voice samples your model needs.
Search and filter your audio by emotional expression, speaking style, vocal qualities, and audio characteristics. Each sample includes transcripts, speaker labels, and measurements across 600+ expression and acoustic dimensions, with links back to your original recordings.
312 samples
- 00:41 · Spk A
“No, that’s actually hilarious.”
Amusement · 0.81Low pitch00:45 · Spk BSource · recordings/rec_0412.wav“I thought you’d like that one.”
ConversationalClean
Mine conversations. Refine your dataset.
Extract use cases and run error analysis across your conversations, with a connection to Hume’s evaluation platform to generate evaluation suites. Filter out low-quality samples, recordings containing PII, and corrupted files to refine the data you use.
- Billing questions3,860
- Order status2,910
Error pattern:Interrupts caller
- Account access1,740
- Returns980
Error pattern:Repeats question
- Passes quality checks9,490Retained
- Low quality1,120Excluded
- Corrupted files214Excluded
- Contains PII806Scrub · optional
Find gaps in data quality and coverage.
Use dataset analytics to see which characteristics are well represented, where examples are missing, and where audio quality needs attention. Identify what to collect, replace, or improve before the next stage of training.
- SNR
- 24 dB
- Clipping
- 7.8%Needs attention
- Transcript WER
- 4.1%
- English58%
- Spanish27%
- Japanese6%Underrepresented
- Calmness41%
- Amusement22%
- Conversational64%
- Read speech36%
- Low pitch33%
- Breathy12%
- Clean71%
- Far-field18%
Why Hume
A research lab’s expertise, built into your pipeline.
Over a decade of expression research.
Hume’s proprietary expression models bring research into how people express and perceive emotion to audio processing at batch scale.
Quality checks at every step.
Preconfigured model choices, language-specific transcription models, phoneme-level verification, and error-rate checks help you build a reliable corpus from messy audio.
Secure and private.
Run Prism in your own AWS, Google Cloud, or Azure environment, so your audio and production data never leave your environment.
Turn your audio into your next training or evaluation dataset.
Talk to Hume about your recordings, development goals, and cloud environment. We’ll help define the processing steps and deployment that fit your team’s workflow.