Silencio-2 puts sighs and whispers in a 112 MB voice model

A computer can finally sigh on cue. Whether the sigh has quite the right amount of exasperation takes another listen. BlueRose Labs’ Silencio-2 accepts stage directions inside the text, giving a small speech model something to perform beyond a string of words.

The model card shows bracketed sounds and tagged delivery changes. A sigh belongs where its marker appears; an emotion applies to the following text until another tag changes it. That is a useful distinction for a creator trying to shape a line. A noise is an event. A whispered sentence is a performance.

Silencio-2 derives from Kyutai’s Pocket TTS. Its weights fit in one 112 MB file, and it produces 24 kHz mono English audio. The card describes added tokenizer tokens and a final distillation stage that reduces generation to one pass per frame. The practical attraction is a small local system with explicit direction, rather than a narrator who needs a new paragraph of coaching for every sound.

BlueRose Labs reports the M4 speed figure in the card. Treating that number as a promise for every laptop would give the benchmark a second career in fiction. A different CPU, configuration or workload needs its own timing, particularly if the goal is live playback while another application is busy.

Voice prompts can come from a short clean recording, and bundled voices offer another starting point. A useful first experiment would use one of those bundled voices and a few short lines. That separates the question of whether the tags work for your script from the extra question of whether your reference recording is any good.

The limitations section does some actual work. BlueRose Labs says, “The delivery changes from take to take,” and calls whispering the least steady tag. A tag can be forceful in one result and faint in another. Saving the seed alongside an accepted take gives a creator a better chance of revisiting a result deliberately, even though listening remains necessary.

There is a second catch hiding in plain hearing. The nonstreaming generate path raises quiet takes toward speech level by default, while streaming leaves the level alone. A whisper’s character and its volume are different things. If a quiet line suddenly seems louder in the export, changing adjectives in the script may be fixing the wrong part of the pipeline.

A local workflow becomes attractive when a creator needs several attempts at the same passage and wants to compare them without adding a network round trip to every revision. The real test is the saved audio: whether the direction is audible, whether the words stay clear, and whether the line fits the scene around it.

The card requires permission for cloned voices and warns against deceptive impersonation. That leaves plenty of room for useful work with an authorized voice. It also leaves the actor in the loop, even when the actor happens to be a very small file.

Even the laptop needs a breather.