Zonos v0.1

State of the art text-to-speech model [model]. [blog], [Zyphra Audio (hosted service)]

Unleashed

Use this space to generate long-form speech up to around ~2 minutes in length. To generate an unlimited length, clone this space and run it locally.

Tips

When providing prefix audio, include the text of the prefix audio in your speech text to ensure a smooth transition.
The appropriate range of Speaking Rate and Pitch STD are highly dependent on the speaker audio. Start with the defaults and adjust as needed.
Emotion sliders do not completely function intuitively, and require some experimentation to get the desired effect.

Zonos Model Variant

Language

Enhance Output with DeepFilterNet

Speech Text (Up to 2000 Characters)

Generated Audio

Long-Form Parameters

Max Chunk Length (Characters)

The maximum number of characters to generate in a single chunk. Zonos itself has a much higher limit than this, but consistency breaks down as you go past ~200 characters or so.

1 300

Target RMS

The target RMS (root-mean-square) amplitude for the generated audio. Each chunk will have its loudness normalized to this value to ensure consistent volume levels.

0 1

Punctuation Pause Duration (Seconds)

Pause duration to add after a chunk that ends with punctuation. Full-stop punctuation (periods) will have the entire length, while shorter pauses will use half of this duration.

0 1

Chunk Cross-Fade Duration (Seconds)

The duration of the cross-fade between chunks. This helps to smooth out transitions between chunks. In general, this should be set to a value greater than the pause duration.

0 1

Generation Parameters

Optional Prefix Audio (continue from this audio)

Equalize Prefix Audio

Enhance Prefix Audio with DeepFilterNet

CFG Scale

1 5

Min P

0 1

Seed

Randomize Seed

Conditioning Parameters

All of these types of conditioning are optional and can be disabled.

Skip Speaker

'Speaker Noised' is a conditioning value that the model understands, not a processing step. Check this box if your input audio is noisy.

Speaker Noised

Equalize Speaker Audio

Enhance Speaker Audio with DeepFilterNet

Speaker Pitch Shift (Cents)

A pitch shift to apply to speaker audio before extracting embeddings. A slight down-shift of ~45 cents tends to produce a more accurate voice cloning.

-1200 1200

Optional Speaker Audio (for cloning)

Skip Emotion

Happiness

0 1

Sadness

0 1

Disgust

0 1

Fear

0 1

Surprise

0 1

Anger

0 1

Other

0 1

Neutral

0 1

Skip DNSMOS

Deep Noise Suppression Mean Opinion Score [arXiv 2010.15258]

1 5

Skip VQScore

VQScore [arXiv 2402.16321]

0.5 0.8

Skip Fmax

Fmax (Hz)

0 22050

Skip Pitch

Pitch Standard Deviation

0 300

Skip Speaking Rate

Speaking Rate

5 30