Why VoxaTrace?
The Gap in Voice AI
Voice AI has made incredible progress understanding what you say. Speech-to-text is solved. But none of it understands how you sound.
Pitch. Timbre. Melody. Rhythm. Emotion. Vocal quality. These acoustic qualities are invisible to speech AI — but they're everything for singing apps, vocal training, music education, and any application where the sound of a voice matters.
VoxaTrace fills that gap.
What Makes VoxaTrace Different
Acoustic Intelligence, Not Speech Recognition
| Feature | Speech AI | VoxaTrace |
|---|---|---|
| Output | Text transcript | Pitch (Hz), confidence, vocal activity |
| Focus | Words and language | How the voice sounds |
| Use case | Dictation, commands | Singing, training, games |
| Latency | 100ms–1s (cloud) | ~64 ms (on-device, BALANCED) |
On-Device, Real-Time
No server calls. No network latency. No data leaving the device.
- ~64 ms latency — fast enough for real-time feedback (BALANCED preset, 1024-sample window at 16 kHz)
- Works offline — no internet required
- Privacy by design — audio never leaves the device
Battle-Tested at Scale
- 8 years of R&D — not a weekend project
- 5 million users — proven in production
- Same API on Android and iOS — write once, ship everywhere
Beyond Mobile (3.0.0)
As of 3.0.0, VoxaTrace also ships a desktop/server JVM target (macOS Apple Silicon and Linux x64) so you can run the same analysis off-device — for example batch lesson authoring on the server. See the JVM Quickstart.
Three-Tier API Design
Use as much complexity as you need:
// Tier 1: Just works (80% of users)
val detector = PitchDetection.createDetector()
// Tier 2: Discoverable options (15% of users)
val detector = PitchDetection.createDetector(
PitchDetectorConfig.Builder()
.algorithm(PitchAlgorithm.SWIFT_F0)
.voiceType(VoiceType.westernSoprano)
.build()
)
// Tier 3: Full control (5% of users)
val config = PitchDetectorConfig.PRECISE.copy(
confidenceThreshold = 0.6f,
amplitudeGateDb = -50f
)
Algorithm Performance
Two pitch detection algorithms for different needs:
| Metric | YIN | SwiftF0 |
|---|---|---|
| Latency | ~50ms | ~50ms |
| Accuracy (clean audio) | Good | Excellent |
| Accuracy (noisy audio) | Fair | Good |
| Memory overhead | Minimal | ~10MB (model) |
| Battery impact | Low | Low-Medium |
| Dependencies | None | ONNX Runtime |
When to Use Each
Choose YIN when:
- Building a simple tuner
- Targeting low-end devices
- Audio is clean and controlled
- You want minimal dependencies
Choose SwiftF0 when:
- Building singing or vocal training apps
- Users are in noisy environments
- Accuracy matters more than app size
- You need robust vibrato detection
What You Can Build
VoxaTrace enables any application where the acoustic qualities of voice matter:
| Application | What VoxaTrace Provides |
|---|---|
| Singing apps | Pitch detection, scoring, performance feedback |
| Vocal training | Real-time intonation analysis, progress tracking |
| Music education | Ear training, sight-singing, pitch matching |
| Voice games | Pitch as game input (sing to jump, hum to control) |
| Accessibility | Voice-based input beyond speech recognition |
| Wellness | Vocal health monitoring, breathing exercises |
Ready to Start?
- Installation — Add VoxaTrace to your project
- Android Quickstart — First app in 5 minutes
- iOS Quickstart — First app in 5 minutes
- Karaoke App Recipe — Complete example with scoring