Voice Cloning Best Practices | Smallest AI Docs

High-quality reference audio is the single most important factor in clone quality. These guidelines cover recording environment, speaking style, multi-lingual cloning, and expressive control.

Try Voice Cloning

Clone a voice directly in the console. 5-15 seconds of audio, no code required.

Recording Reference Audio

Environment

Record in a quiet room with minimal background noise. Ambient noise, hiss, or rumble will be captured in the clone.
Use a dedicated microphone when possible. MacBook and mobile device microphones are acceptable if positioned at an appropriate distance to avoid distortion.
Avoid rooms with echo (large empty spaces, outdoor areas). Small treated rooms produce the best results.
After recording, listen back to the audio before uploading. Verify it is free of interruptions, clipping, or background interference.

Speaking Style

Speak naturally in your normal conversational voice. The model captures timbre, accent, emotional tone, rhythm, and pacing automatically.
Maintain a consistent pace throughout the recording. Avoid long pauses, as they can degrade clone quality.
Do not exaggerate emotion unless a specific tone is the intended output (see Expressive Cloning below).

Audio Length

Provide 5 to 15 seconds of clean, continuous speech.

Multi-Lingual Voice Cloning

Language Matching

For best results, record reference audio in the same language as your intended output. The model supports cross-lingual cloning (e.g., English reference audio used for Spanish output), but a language-matched reference will always produce higher fidelity.

Scenario	Expected Quality
Reference and output in the same language	Best results. Highest phonetic accuracy.
Reference in a different language than output	Functional. Voice characteristics transfer, but the source accent is retained.

Accent Retention

When synthesizing in a different language than the reference audio, the original accent is preserved. A clone from a South Indian English speaker will retain that accent when generating Hindi or Tamil output. This is by design: the clone reproduces your voice, including accent characteristics.

If accent-neutral output is required for a specific language, provide reference audio recorded by a native speaker of that language.

Language Group Constraints

Cloned voices follow the same language group routing rules as standard synthesis. See Code-Switching for details on Indic and Global group restrictions.

Expressive Cloning

The model captures emotional and prosodic characteristics from the reference audio. The tone, pace, and volume of the reference directly influence the synthesized output.

Emotional Control

The emotion conveyed in the reference audio (e.g., calm, happy, angry) is reflected in the generated speech. To produce an angry-sounding clone, provide an angry reference. To produce a neutral clone, provide a neutral reference.

Speed Control

The pace of the reference audio determines the output speed. A fast-paced reference produces faster delivery; a slower reference produces more measured output.

Volume Control

The volume level in the reference audio carries over to the output. A soft-spoken reference produces quieter output; a louder, more energetic recording produces bolder output.