Voice Studio in Depth

Voice Studio is six separate tools on one ElevenLabs account. They share one credit pool, and the hub page says so in a red warning box: a few minutes of voice generation can eat a month's quota. So the first skill is spending credits on the takes that matter and none on the ones you could have predicted.

Text2Speech: the settings that do the work

Text2Speech has two modes, Voice Only and Voiceover on Video. Start with Voice Only. You pick a voice from the sidebar (filters for gender, accent, age and use case, plus search), paste the script, set the tuning, and generate. The voice list shows voices with preview clips first and caps results at 100, so filter rather than scroll.

The last row needs a warning. The help page explains that pitch is changed after generation using FFmpeg's asetrate and atempo filters. That method moves the voice's formants, the resonances that make a throat sound like a throat, along with the pitch. SoundStudio's theory page on time and pitch describes the result: the chipmunk sound. Two semitones is usually safe. Five is a cartoon. If you need a lower voice, pick a lower voice.

Speed has its own ceiling for a different reason. ElevenLabs' documentation says extreme values may hurt quality. A script that only fits at 1.20x is a script that needs cutting.

Models and what they cost

The sensible pattern is to draft on Turbo or Flash and render the final on Multilingual v2. One trap: ElevenLabs' docs warn that Flash v2.5 is weaker at reading numbers and currency. Draft takes on Flash may mispronounce figures the final model reads correctly, so do not "fix" a number in the script until you hear it on the final model.

One more limit to know. The Text2Speech script box counts up to 10,000 characters, while the help page says 5,000 per generation. When two pages disagree, use the lower number: generate in sections of under 5,000 characters, split at paragraph breaks, and join them in SoundStudio.

Captions and video voiceover

With captions set to Auto-generate, ElevenLabs returns word-level timestamps and the tool builds an SRT with lines of about 4 to 8 words, breaking at punctuation. The help page says this costs no extra characters. You can download SRT, VTT or TXT. If you paste your own SRT instead, it is not re-timed to the new audio.

Voiceover on Video takes an MP4, MOV or WEBM and offers three mixes: replace the original audio, mix with it (the voiceover at +80 percent gain), or keep the original and add the voiceover after it. "Burn captions" draws the words onto the frames. This all runs in your browser with FFmpeg.wasm, so the help page suggests videos under 200 MB and a desktop browser. For anything with music or several clips, you will get better control generating Voice Only audio and assembling in VideoStudio or AVPrep3.

Output is MP3 at 192 kbps by default, 16-bit WAV, or OGG Vorbis. After a generation, Convert to buttons re-encode to another format without regenerating, so there is no credit cost. Export WAV for editing. Always.

Speech2Text: the proofreader you did not know you had

Speech2Text transcribes audio or video (MP3, WAV, MP4, MOV, WEBM, M4A, OGG, FLAC) with the Scribe v1 model. Options include auto-detect or a fixed language, word or character granularity, diarization to label speakers, and tagging audio events like laughter or music. Exports: TXT, SRT, VTT, JSON with all timings, and CSV with one row per utterance. The hub says it supports 99 languages.

Its obvious job is transcribing interviews. Its sneaky job is checking synthetic narration. Codec-style speech models occasionally drop a word, repeat one or say something slightly different. Your ear slides over that on the fifth listen. A transcript does not. Run the final narration through Speech2Text, paste the transcript next to your script, and read them side by side. Any mismatch is a line to regenerate or a pronunciation to fix. Transcription costs credits too, so do it once, on the final.

The other three tools

SoundEffects prompts work like a sound designer's notes. The page's tips ask for the object and environment, intensity, atmosphere, motion and materials. "Footsteps" gives you something. "Slow boots on a wooden porch, close, dry, a board creaks on the third step" gives you a scene.

AudioIsolation is strong and honest about its limits. The page notes that very loud noise may leak through, whispers may soften and multiple speakers are preserved. If it is used on an interview that gets posted, the voice was processed by AI, and the edit should not change meaning. Cleaning noise is fine. Do not use it to make a phone call sound like a studio interview without saying it was a phone call.

MusicGeneration's page says ElevenLabs Music supports 10 seconds to 5 minutes per generation. ElevenLabs' own docs, as of October 2026, give a minimum of 3 seconds and a maximum of 5 minutes. The tool's duration slider starts at 10 seconds, so the difference rarely matters. Lesson 5 explains how these music models work, and lesson 6 covers the deeper song tools in MusicStudio.