Voice Playground
Safely test how your Delphi sounds before you change global settings.
What it is
Voice Playground is a sandbox for trying voices, scripts, and settings without affecting live calls. Adjust sliders, switch models, and audition pronunciations—then apply only when you’re happy.
Where: Studio → Identity → Voice → Playground (Generate tab)
Why use it
Risk‑free tuning: Experiment freely without changing how callers hear you.
Faster iteration: Swap scripts, tweak sliders, compare models side‑by‑side.
Precision fixes: Dial in accents, pacing, and tricky words before going live.
Quick Start (2 minutes)
Open Voice → Open Voice Playground and switch to the Generate tab.
Paste or type a short test script (what your Delphi would actually say on a call).
Adjust Stability / Similarity / Speed; try a Voice Model.
Click Generate to preview.
Like it? Click Apply settings to my Delphi to make them global. If not, tweak and regenerate.
Tip: The truest preview is a live call. After you apply settings, place a quick call and listen end‑to‑end.
The controls (what they do)
Stability
(0 – 100 %)
Highly animated voice—big swings in pitch, loud/soft, intense, whispers.
Locked-in broadcaster tone—more monotone, same volume throughout.
Long reads that need consistency.
Role-play or emotional storytelling.
Similarity
(0 – 100 %)
Studio-polished synthesis—background hiss removed, quirks smoothed out.
Carbon copy of your raw sample—every breath, accent edge, and mic artifact preserved.
Brand voice that must sound exactly like you.
Noisy or low-quality original sample.
Speed
(0.7× – 1.2×)
0.7× = 30 % slower for clarity.
1.2× = 20 % faster for snappy updates.
Quick Q&A sessions.
Dense explanations, language learners.
Quick rules of thumb
Stability: Drop by 10 % if calls feel flat; raise if voice sounds chaotic.
Similarity: Keep below 60 % unless your sample was studio-grade quiet.
Speed: Adjust in 0.05 steps to avoid sounding rushed or sluggish.
Click Reset Settings anytime for a clean slate: 50% stability, 75% similarity, 1x speed.
Voice Models
Default — Balanced clarity; softens minor accent edges.
For Accents 1 — Most faithful accent reproduction (slower than Accents 2).
For Accents 2 — Fastest of the accent models; keeps general flavor, may miss fine nuances.
Workflow: Switch model first, then nudge Stability/Similarity in 10‑point steps.
Custom Pronunciations
Custom pronunciations only work in Voice Model - Default
Click Settings Icon (top right) → Scroll down.
Click Open next to Custom Pronunciations
Word: e.g., Delphi.
Phonetic: e.g., DEL‑f‑eye.
Add and preview in Playground; tweak spelling if it sounds off.
Remove with the trash icon.
Sample scripts (copy → paste)
Intro (20–30s): “Hi, I’m [Name]. I can help with [two topics]. What are you working on this week?”
Explainer (40–60s): “Here’s a quick plan: first, define the goal; second, pick one channel; third, test with a small cohort…”
Coaching tone check: “You’re closer than you think. Let’s try that sentence again—this time focus on the verb order.”
Best practices
Record once, tune many: Start with one clean 30-sec sample; use Playground for the rest.
Match reality: Test with real phrases you’ll use (intros, FAQs, tough names).
One variable at a time: Change a slider or model, generate, then compare.
Mind the room: If previews sound off, the sample may be noisy—re‑record rather than over‑tuning.
Confirm live: Always validate changes with a short real call.
FAQs
Do Playground changes affect callers? Not until you click Apply settings to my Delphi.
My accent isn’t captured. Try For Accents 1 first; if latency matters, use Accents 2 and increase Similarity slightly.
How do I get it to pronounce names correctly? To get your Delphi to pronounce names correctly, use custom pronunciations. Add the word as you want it to be said phonetically.
Why does my voice sound different on a live call versus the Playground or Read Aloud? Playground clips are “experimental,” rendered offline and a bit slower, so they can use looser settings that change each time. Calls must stream in real time, so Delphi adds extra stability for speed and consistency. To hear the truest result, tweak settings and then start a live call. Use Playground only for creative reads, ads, or long scripts.
Why does the Playground and the Read Aloud function generate a slightly different take every time? The reason the Playground and Read Aloud functions generate a slightly different take every time, even if the settings are the same, is that both run on generative AI text-to-speech technology. Each time you click Generate or Read Aloud, Delphi samples tiny shifts in pitch, timing, and energy—like rolling fresh dice inside the same rules. That touch of randomness keeps the voice from sounding canned, but it also means no two clips are identical. To widen the picture, this is how generative AI works for all AI platforms! It's the reason that ChatGPT won't give you the same answer to one question asked twice.
Troubleshooting
Choppy preview: Reduce Speed slightly and increase Stability ~5–10 pts.
Too robotic: Raise Similarity 5–10 pts; read a better sample.
Mushy diction: Lower Similarity a bit; re‑record with a closer mic.
Names mispronounced: Add Custom Pronunciations; test multiple phonetic spellings.
Before you apply
Last updated