Word-level voice editing with Tonality from Riverside Labs

Individual words, pauses, and intonation can be edited with Tonality. Riverside Labs offers a free playground and API access through sales.

Replacing a name in an approved take can mean rerecording a sentence, then trying to match its rhythm and the original recording conditions. Tonality from Riverside Labs targets that correction process: changing words in a recording and how they are delivered while preserving, according to the company, the speaker’s voice and the original room acoustics.

The announcement introduces the work of Riverside Labs, Riverside’s new audio and video research lab. The service works with existing recordings, but it can also generate new speech from text and a reference voice. Its focus is on precise changes, down to a single word.

The interface supports two kinds of edits. Users can replace, insert, or delete words in the transcript, as well as add sentences. Separate controls adjust pitch, tone, duration, pauses, energy, and pronunciation. An edit can therefore change what is said or how it is delivered, without necessarily rewriting the passage.

The workflow begins with uploading a file or entering a script. Users then select the words they want to change, adjust their delivery, and lock the sections they want to keep. After generation, the service returns the edited audio and the exact sample ranges it reports as changed. That information is intended to help users locate the edits and check their extent.

Preserving the rest of the take is the product’s central promise. That encompasses several distinct qualities: maintaining the speaker’s identity, accent, and pacing, retaining the sound of the room, and preserving untouched passages. The published examples illustrate the intended result; they do not establish that those qualities will hold for every voice and recording.

An audio comparison on the product page applies the same script changes across several services, including HeyGen, Cartesia, ElevenLabs, Qwen, and Fish Audio. The edits include changing four short videos to six and adding a few words elsewhere. These are examples selected by Riverside, without a detailed evaluation protocol accompanying the presentation. On their own, they cannot establish a ranking across the full range of speech-editing tasks.

The claim of being the first also needs qualification. Research such as VoiceCraft, published in 2024, already addressed inserting, deleting, and replacing words in recorded speech. That precedent does not establish equivalence with the six controls offered here. It does show that localized speech editing predates Tonality. Riverside’s claimed distinction lies in combining those controls while preserving the surrounding recording.

For production teams, potential uses include corrections after approval: updating an offer in a podcast ad, replacing a name, or correcting pronunciation in an audiobook passage. In dubbing, duration and pause controls could help fit a line into the available time. Changes in duration still need to be checked in the edit; the presentation does not explain how any resulting timing shift is compensated for.

The free playground provides access to the controls on users’ own recordings, subject to usage limits. API access is available through the sales team, which provides keys, documentation, and volume-based pricing terms. No public per-unit price accompanies the offer. Teams planning to integrate it into an automated production workflow will therefore need those terms before estimating the cost of their edits.