Trast Sales

How to make the model 1000% accurate

Training a custom voice for Riva is fundamentally a data-quality problem, not a compute problem. A clean 30 minutes outperforms a sloppy 5 hours. Read this whole page before recording — most issues are unrecoverable after capture.

1. The room

• Closet method (best): sit inside a clothes closet. Clothes absorb echo. Best free studio you have.
• Soft surfaces beat hard ones. Carpets, curtains, sofa, bed — all absorb. Glass, tile, drywall — all reflect.
• Avoid bathrooms, kitchens, empty rooms. Reverb leaks into the recording and becomes a permanent artifact in the cloned voice.
• Kill all noise sources: HVAC off, fridge off if possible, phones on silent, dogs out. Even a faint fan-hum is learnt by the model.
• Same room every session. Recording two sessions in different rooms produces audible "voice changes" in the synth. Pick one space, commit to it.

2. The microphone

• Required class: USB condenser, not the built-in laptop mic. Built-in mics record room more than voice.
• Recommended: RØDE NT-USB+ (€200), Shure MV7 (€250), Blue Yeti X (€170). All are fine.
• Pop filter — non-negotiable. €10. Without one, every "P" and "B" peaks and clips. Clipping is unrecoverable.
• Distance: 15-20 cm with the pop filter between you and the capsule. Same distance every session.
• Cardioid polar pattern. Most USB condensers default to this; check your mic's settings.

3. Your voice

• Same time of day. Voice changes through the day (morning gravelly, evening tired). Pick one slot, stick to it. Mid-morning is generally most consistent.
• Hydrate. Water before and during. Avoid dairy (creates phlegm). Avoid coffee right before (dries out vocal cords).
• Sit, don't stand. More consistent posture and breath support.
• 30-min sessions max. Take 10-min breaks. Voice fatigue produces inconsistent training data — a worse problem than less data.
• Read in your normal speaking voice, not a "presenter" voice. The bot will sound like the voice you train. Demo-acceptance favours conversational tone.
• Pace yourself. Don't rush. Let punctuation breathe. Comma → small pause. Period → full pause. Question → rising pitch.

4. Dataset size — find your minimum

Don't commit to 5 hours upfront. Record 30 min, train, listen. Decide if quality is acceptable; if not, record more. The kit's "Record" tab shows your running net-audio total per stage.

5. Phonetic coverage matters more than raw hours

5 hours of "the cat sat on the mat" repeated produces a model that knows only those phonemes. Your prompt list must cover:

• Every vowel in stressed and unstressed positions.
• Every consonant in initial, medial, and final positions of words.
• All diphthongs (Greek: αι, οι, ει, αυ, ευ).
• Tricky consonant clusters (Greek: στ, στρ, τζ, τσ, μπ, ντ, γκ, νθ, χθ).
• All sentence intonations — declarative, interrogative, exclamatory. Without question prompts you'll get a model that can't ask questions.
• Numbers, dates, currency spelled out at least 50 times each.

The included stage_0_starter.txt is a curated 50-prompt set covering common Greek casino phonetics. For longer datasets, use Mozilla Common Voice Greek's prompt list — it's already phonetically balanced by construction.

6. Pre-flight checklist

1. Quiet room with soft surfaces ✓
2. Mic at fixed 15-20 cm distance with pop filter ✓
3. Mic level peaks at ~70-90% on loudest words (no clipping) ✓
4. Test recording one prompt → listen back → no echo, no hum, no breath blast ✓
5. Same time-of-day slot scheduled for all sessions ✓

7. Common pitfalls

• Mouth noises (clicks, smacks) — swallow before each prompt. Lip-smacks are the hardest artifact to remove.
• Inconsistent distance — voice volume changes proportional to distance². Mark a fixed spot.
• Re-recording dirty takes — never keep a take with a mistake "to fix later". Press redo, do it clean. The model learns whatever is in the data.
• Reading in a robot voice — your speaking voice is what the bot will sound like. Speak naturally as if telling a friend.
• Recording when sick/congested — your voice is different. Skip that day, resume when healthy.

8. After recording — what we do

1. Quality QC: trim silence, detect clipping, check loudness consistency, listen to random 5% of clips.
2. Forced alignment with Montreal Forced Aligner — produces phoneme-level timestamps NeMo needs.
3. FastPitch training on a separate L40S/A100 GPU — 18-24 hours.
4. HiFi-GAN fine-tune from NVIDIA's universal vocoder — 3-6 hours.
5. Convert to Riva via nemo2riva + riva-build.
6. Deploy as a new riva-speech-greek service on the Nebius box (separate gRPC port from Magpie).
7. You test; if not good enough, record more, retrain.

9. The 1000%-accuracy reality check

Voice cloning quality plateaus around what the dataset supports. With our pipeline:

• 5 h clean dataset + good phonetic balance + clean room: ~95% intelligibility, mostly natural prosody. Sounds like you.
• 20 h with a voice actor and studio recording: ~99% intelligibility, indistinguishable in casual listening.
• "Indistinguishable from real human in adversarial testing" requires either hundreds of hours or model architectures (e.g. VALL-E-X scale) that don't exist self-hosted yet.

"1000% accurate" doesn't exist. But "indistinguishable for a 2-min sales call" is achievable with 5 hours and discipline.