USER GUIDE

Voice Studio end to end

Every image on this page is a live screenshot of the running app. Voice Studio works entirely offline — the models run on your machine and no audio is ever sent anywhere.

The main G-Labs Voice Studio window: seven-item sidebar and the Voice Clone tab
The main window. The left sidebar holds the seven working screens; below them sit Detailed Logs, Settings and your plan status.

What Voice Studio is

A desktop app that runs speech models directly on your machine. No internet needed to generate audio, no per-character fees, and the audio never leaves your computer.

🔊

Generate voices

Clone a voice from a short sample, or build a new one from voice attributes.

💬

Many characters

Dialogue scripts where every character gets their own voice.

📝

Subtitles

Transcribe audio/video, then translate the subtitles with an LLM.

Modest machines still work: the app checks your GPU on its own; if the card is not compatible it switches to multi-core CPU and says so clearly, instead of crashing with a cryptic CUDA error.

Install & load models

The first run needs to download the models. Go to Model Manager in the sidebar, download the speech (TTS) package — that one is required — and at least one recognition model if you plan to extract subtitles.

The Model Manager screen: TTS model status and the recognition models with their sizes
Every model lists its size and the VRAM it needs. Models you already have show ✓ Downloaded; the rest have a Download button.
Keep the models on an SSD. The app loads them into RAM/VRAM at startup — on an HDD every launch means a long wait.
SCREEN 01

Voice Clone

Clone a voice from 5–10 seconds of reference audio. This is the screen the app opens on.

  • Reference audio file — drag and drop or browse for a file. The cleaner the sample (no background music, no overlapping speakers), the closer the clone.
  • Reference text (required) — type out exactly what is said in the sample. A single wrong word skews the voice. If you would rather not type it, hit AI suggest and Whisper fills it in.
  • Text content — enter the lines you want read in the cloned voice, one sentence per line.
The Voice Clone tab: reference file picker, reference text, advanced settings and the Voice Library
The sample and its settings sit on the left; the table of lines to read and the preview player are on the right.

Advanced Settings

Detail level (number of generation steps), Guidance strength (how tightly it follows the sample) and Speaking speed. Every field has a ↺ button to restore the default — experiment freely, nothing breaks.

SCREEN 02

Read Text

Generate a voice with no reference audio at all. Pick a preset voice from the library, or hit random voice and let the app roll a fresh one.

The Read Text tab: preset or random voice selection, the content table and the settings groups
Pick a voice from the library, then paste your script into the table — ideal when you have a pile of lines to get through.
This is the natural screen for Batch Processing: one fixed voice, many lines, exported as separate files or merged into one file with a matching .srt.
SCREEN 03

Group Dialogue

Multi-character scripts, one voice per speaker. Write in the <Name>: line of dialogue format, then assign a voice to each name in the voice casting panel.

The Group Dialogue tab: script editor, voice casting panel and the generation queue
There is a sample dialogue button that shows the exact format, and an analyze dialogue button that pulls the characters out for you.

Great for podcasts, audio drama and interview videos — anything where a single flat narrator gets boring fast.

SCREEN 04

Extract Subtitles

Turn speech in audio or video (MP3, WAV, M4A, FLAC, MP4, MOV…) into text with timecodes.

The Extract Subtitles tab: file picker, recognition model selector and the extracted results table
Pick the recognition model to match your machine: Turbo is fast and good enough for most jobs; Large v3 is the most accurate but the heaviest.
  • Audio language — "Auto-detect" works fine, but naming the language explicitly is more accurate on short clips.
  • Export plain .txt, or .srt with timecodes to drop straight into your video.
  • One button sends the result over to the Translate tab.
SCREEN 05

Translate

Translate a subtitle file with an LLM without breaking the timing: the line count and the timecodes are held absolutely fixed, and the AI is only allowed to change the words.

The Translate tab: file input, target language, the consistency table and the translated results table
Accepts .srt, .vtt, .ass, .sbv and plain .txt. You can also pull the result straight from the Extract Subtitles tab.

Consistency table

Before translating, the app reads the whole file to lock in the tone, terminology and forms of address, then feeds that table into every chunk it translates. This is what stops proper nouns and pronouns from drifting halfway through — and you can edit the table before the run starts.

You need to pick an LLM first under Model Manager. Claude CLI, Antigravity, Codex or the 9Router gateway all work.
SCREEN 06

Model Manager

One place for everything heavy: the speech model, the recognition models, and the LLM used for subtitle translation.

The Model Manager screen: AI models plus the LLM area with 9Router, Claude CLI, Antigravity and Codex
The LLM area sits below: each provider has its own status line and a Guide button with the install command ready to copy.
  • 9Router — a gateway running on your machine that funnels several model sources into a single endpoint.
  • Claude CLI · Antigravity · Codex — use the CLI plan you already pay for, no separate API key needed.
  • Hit Refresh after installing a CLI so the app rescans and repopulates the model list.
SCREEN 07

Webhook API

Turn on a local HTTP server so other tools can call in to generate speech — n8n, Make, Zapier, a Python script, or G-Labs Story Machine.

The Webhook API screen: server controls, API key and usage instructions with a JSON example
Start the server, copy the API key and the URL — the instructions, a sample JSON body and a cURL example are right below.
This is exactly how Story Machine gets its narration: start the server on port 8766, paste the URL and the API key into Story Machine, done.

Voice Library

Shared across all three audio-generating screens. The app ships with 30 preset voices (male/female, a range of timbres); the voices you clone are saved here with a name and a short description.

Pin favorites

Hit ★ to push the voices you use most to the top of the list.

Move to another machine

Back up and restore with a .vcp file — you can share it with other people too.

Tips & troubleshooting

The clone does not sound like the sample

Nine times out of ten the reference text does not match the audio. Listen again and type it word for word, or press AI suggest. Background music or two overlapping speakers will skew the voice as well — cut out the cleanest 5–10 seconds.

The app misreads special characters (100%, 25°C, m²…)

Use the pronunciation dictionary: spell it out phonetically once and the app remembers it per language and applies it from then on.

Can I use it without an NVIDIA GPU

Yes. The app falls back to multi-core CPU (and MPS on Apple Silicon). Slower, but it works.

Subtitle extraction misspells proper nouns

Switch to a bigger model (Large v3) and set the audio language explicitly instead of leaving it on auto-detect.

Subtitle translation throws the line count off

It cannot happen — the app hard-locks the line count and the timing. If one line fails to translate, it keeps the original text and flags it with ⚠ for you to fix by hand, rather than desyncing the whole file.

Story Machine says Voice Studio is not configured

Open the Webhook API screen, press Start on the server, then copy the URL and API key into Story Machine's settings.