ElevenLabs

Give your assistant the text and it can list the voices on your account, check which models are available and what each one can do, then produce a downloadable audio file. The one mechanical rule worth knowing is in the connector's own note: keep the voice, the model, and the output format identical across every chunk of a long document, or the joins between chunks are audible.

What you end up with. An audio file of the document you can play on a commute, and a history entry you can pull the file from again later.

Start with it doing the work and checking with you.

Never without. Check what the text costs in characters before a long document goes through. Generation bills per character, and a hundred-page report is a different order of spend than a memo.

Built on: Get voices list / Get models / Text to speech / Get generated items

Rather than clicking through the library, it can list the voices already on your account with their full attributes, search the shared voice library with filters, and find library voices similar to an audio sample you provide. It can also read the current stability, similarity, style, and speaker boost settings on any voice, and the service defaults those fall back to. Nothing is generated and nothing is changed.

What you end up with. A shortlist of candidate voices with the settings each one is currently configured with, ready to test against your script.

Start with it bringing you the information and you deciding.

Never without. A voice that reads well in a sample sentence can fall apart across ten minutes of your actual copy. Judge the shortlist on a real paragraph, not on the library preview.

Built on: Get voices list / Get shared voices / Get similar library voices / Get voice / Get voice settings / Get default voice settings

It can pull character usage as a time series for you or the whole workspace, broken down by dimension such as voice or user, read the subscription details that allowance sits inside, and list the generation history behind the numbers. The breakdown is the useful part. A total tells you that you are close to the limit, and the breakdown tells you which project got you there.

What you end up with. A usage trend with the consumption attributed by voice and by person, set against what your plan actually includes.

Start with it bringing you the information and you deciding.

Never without. Characters generated is not the same as characters billed. Retries, failed conversions, and streaming behave differently, so treat this as your consumption pattern rather than as the invoice.

Built on: Get Usage Character Stats / Get user subscription info / Get generated items

Pronunciation dictionaries exist for this and they are writable through the connection. It can list the dictionaries you already have, create one from a set of rules, and add more rules as you find them. Rules come in two kinds: a plain alias that swaps one spelling for another, and a phoneme rule written in IPA when the alias is not close enough. It can then attach the dictionary to a project so everything generated afterward uses it.

What you end up with. A named dictionary holding your product names, people's names, and acronyms, attached to the project so the corrections stick.

Start with it doing the work and checking with you.

Never without. Confirm each rule by listening to it, and remember that a dictionary changes only what is generated from now on. Audio you already produced keeps the old pronunciation until you regenerate it.

Built on: Get pronunciation dictionaries / Add pronunciation dictionary from rules / Add rules to the pronunciation dictionary / Update project pronunciation dictionaries

It can create a project and initialize its content from a document or a URL, read back the chapter list with the conversion status of each one, convert a single chapter to check the result, and then convert the whole project using its configured voices and settings. Snapshots are readable too, so you can see the earlier versions of a chapter rather than wondering what changed.

What you end up with. A narrated project with each chapter converted and its status visible, plus the snapshot history behind each one.

Start with it doing the work and checking with you.

Never without. Convert one chapter and listen to it before approving the full project. A whole-project conversion spends the character allowance for every chapter at once, and a pronunciation or pacing problem you would have caught in chapter one is then in all thirty.

Built on: Add new project with attributes / Get projects / Get project by ID / Get chapters by project id / Convert chapter to audio / Convert a project / Get project snapshots

It can submit a video or audio file, or a source URL, for dubbing into a target language, then track the job and collect the results: the project metadata and status, the transcript in the dubbed language, that transcript exported as SRT or WebVTT with timing, and the dubbed audio itself. Subtitles and dubbed audio come out of the same job, which is the part people usually miss and pay for twice.

What you end up with. A dubbed audio track plus a timed subtitle file in the target language, both traceable back to one dubbing project.

Start with it doing the work and checking with you.

Never without. Have someone who speaks the language read the transcript before the dub is published anywhere. Machine dubbing gets product names, legal wording, and idiom wrong in ways the timing and the voice quality will not reveal to you.

Built on: Dub a video or an audio file / List Dubs / Get dubbing project metadata / Get dubbing transcript by language / Get dubbing transcript in specific format / Get dubbed audio for a language

If you run a Conversational AI agent, the conversations are readable. It can list the agents on the account with their metadata, pull conversations filtered by agent, time range, duration, or rating, then fetch any single conversation in full for transcript-level reading. There is also a live count of conversations happening right now. Reviewing the failures weekly is how the agent's prompt gets better.

What you end up with. A themed read of what people asked, where the agent handled it, and the specific transcripts where it did not.

Start with it bringing you the information and you deciding.

Never without. These transcripts are recordings of real people talking to your company. Decide who may read them and what may be quoted out of them before you start pulling them into an assistant, and treat anything a caller volunteered as covered by your own privacy commitments.

Built on: Get ConvAI Agents Summaries / Get Conversational AI Conversations / Get conversation by ID / Get conversational AI analytics live count

It can list the generation history with filters, pull the metadata and the audio for any single item, download the ones worth keeping as a file or a zip, and then delete the rest. The connector's own wording on the delete is permanent and irreversible, and it takes the audio file with it as well as the metadata.

What you end up with. The clutter gone from the generation history, the takes worth keeping downloaded first, and a written list of exactly what was removed.

Start with it doing the work and checking with you.

Never without. Download everything worth keeping before a single delete runs, and approve the items individually rather than by a filter rule. ElevenLabs describes deleting a history item as irreversible and it takes the audio with it, so a take you meant to keep is a re-generation and a different result.

Built on: Get generated items / Get history item by id / Get audio from history item / Download history items / Delete history item