Audio

At Audio you make sound: voice-overs, conversations between up to 10 voices, songs with sections and lyrics, sound effects and voices of your own. Describe what it should sound like and the AI director makes it, or make each sound yourself. Every sound lands on a multi-track editor, where you mix it and export one finished file.

It's paid from your credit, and every sound shows its price before it's made. A sound that fails is free. The voices speak more than 90 languages; the page is in English and Chinese.

StartCopy Anchor Link

New audio opens with one box: "What should it sound like?"

• Describe it. Write a brief, up to 2,000 characters, and the director takes it from there. Under the box: Kind (let the director choose, or Voice-over, Conversation, Song, Sound effect), Length (15 s, 30 s, 1 min, 3 min or Any) and Language. Add up to 3 audio files of 50 MB each, such as a recording to clone or clean up.
• Or make it yourself. Voice-over, Conversation, Song, Sound effect, Fix a recording or New voice opens an empty project on that form.
• From the AI chat. Ask the AI chat for a voice-over, a jingle or a sound effect: it writes the brief and opens Audio with the director on it.

Under the box, four examples show what a brief becomes: press play to hear one, or pick it to start from its words. You can look around, hear the examples and the voices without signing in; making anything asks you to sign in, and your words stay in the box.

A project started from a brief opens with the director, one started by hand on the Make panel. Switch at any time with Talk to the director or Edit by hand. Listening to voices is free: hear them first.

The directorCopy Anchor Link

The director is Claude Opus 5.5, working on your project step by step. For a brief it plans the piece (what it's for, its length, the voices, the mood, the music), finds and casts voices, writes the words, makes the sounds in batches, places each one on its track, mixes them, listens, fixes what it hears and finishes. You see every sound land on the timeline as it's made. Each run is one version you can undo.

Then ask for changes in the same box: a warmer voice, music under it, the same in Chinese. Pick a sound first to ask about that sound alone.

What it spendsCopy Anchor Link

• A run's sounds go ahead without asking while they cost $2 or less in all.
• Past $2, it proposes first: the plan, each item's price and the total, with Go ahead · up to $3.40 and Change the plan. A brief that looks bigger than $2 starts with the plan.
• Once you approve a budget, it spends within it and asks again before passing it.
• Keeping a voice it designed asks you, unless you picked that preview yourself. Cloning a voice always asks.
• Choices come as cards: play each option (free), then Use this.
• A spend you decline is never tried again: the director finds another way or stops.

The director's own work is charged as it goes, at its model's price, like any AI call. Each sound costs what the table says.

The listen checkCopy Anchor Link

Check the mix listens to the project, or to the sounds just made:

• Loudness against your target, peaks, and silences of 1.5 s or more, measured on our server.
• Words: what was said is written down and compared with the script, so a missing or extra word, a name said wrongly or a line cut off at its end shows up.
• AI listening review: an AI model hears a light copy and notes pacing, delivery, pronunciation, balance and glitches, each on its sound.

Each finding links to its moment, most with a fix. The check costs about $0.01 a minute, priced on its button. The director checks after building and again after fixing, at most 4 times and with 2 AI reviews a run. A finding it has tried to fix twice is left for you.

StoppingCopy Anchor Link

Stop ends the run. Sounds still waiting in line are cancelled, free. Sounds already being made can't be stopped: they finish, land as takes marked "Made after the run stopped", and are charged. Leaving the page during a run works the same way. Sounds you make by hand never stop when you leave: they land and wait for you.

One run at a time per account, up to 48 steps each. You can ask 6 times a minute, 120 an hour and 600 a day.

Make it by handCopy Anchor Link

Edit by hand shows two tabs: Make, with a form for each kind of sound, and Edit, for the sound or track you pick.

The Make button carries the price ("Make · $0.03"). Where the length isn't known ahead, it shows the most you pay ("Make · up to $1.20") and you're charged for the real length. Your first Make, and any over $2, asks you to confirm. A Make never blocks the next: 3 sounds are made at once and up to 20 wait in line, each showing its place. A new sound goes at the playhead on its kind's track.

Voice-overCopy Anchor Link

More settings: delivery presets (Narration, Conversation, News, Drama) that set stability and similarity; style and speed on Eleven v3 and Multilingual v2 (speed 0.7 to 1.2, or 0.25 to 4 with the wider range); speaker boost; language; how numbers are read; seed; quality. Eleven v4 takes its delivery from tags instead of style and speed. Flash reads digits as written, so write numbers as words.

Words longer than the model takes in one go are split at sentence ends into parts that carry on from each other, priced together: "Make 3 parts · $0.41".

ConversationCopy Anchor Link

Up to 10 speakers, each with a voice, and their turns in order. Paste lines like "Name: words" and each becomes a turn. Eleven v4 or v3. A request holds 2,000 characters; a longer scene is made in linked parts, priced together. Change voice for all of Mia's lines shows how many lines are made again, and the price, first.

SongCopy Anchor Link

• Describe it: a prompt, and a length from 0:03 to 10:00 (1:00 to start), or let the model choose: then it's held at the longest and charged for the real length. Instrumental on or off. Music v2.5, or Music v2.
• By sections: Make a plan from my description (free), then shape the song map: up to 30 sections of 0:03 to 2:00, each with a label, lyrics (30 lines of up to 200 characters), styles it sounds like and styles to leave out (up to 50 each), and how closely it follows that section. The first section's 6 or 7 styles set the whole song's sound. Styles are sent in English; written in another language, they're translated when you press Make. Lyrics with [Verse] and [Chorus] headers become sections.
• A change after the song is made remakes the whole song with the same seed, so other sections may change a little. Nothing changes until you press Make again.
• Refused prompts: a prompt that names an artist or a song is refused, not charged, and comes back with a rewrite to Use it or Edit.
• Split into stems (Edit): 2 (vocals and music) or 6. The stems go on tracks under the song, which is muted, not deleted.
• Songs carry Content Credentials, a signature saying they were made with AI.

Sound effectCopy Anchor Link

Describe the sound (up to 450 characters) and set its length: Auto, or 0.5 to 30 s. Loop makes a background that repeats and goes on Atmosphere. Follows the words runs from 0 (more freedom) to 1 (closer to the words), 0.3 to start. 2 takes to start, up to 4: a 5 s effect costs under a cent. Longer than 30 s is made as a 30 s loop and repeated.

Fix a recordingCopy Anchor Link

Upload a file (MP3, WAV, M4A, FLAC or OGG, up to 50 MB) or record one, then:

• Change voice: the same performance in another voice, on Multilingual STS v2 or English STS v2. Up to 5 minutes a part; longer recordings are made in parts and joined.
• Clean up: the background noise taken out. Cleaned amount (0 to 100%) blends the original back under it, free, if the speech sounds robotic.

Both make a new take of the same sound, so the original stays.

Takes and compareCopy Anchor Link

Make always makes a new sound. Another take, Change voice, Clean up and Remake make new takes of the same sound, so nothing is lost. The takes list shows each take's price and what differs ("Stability 0.7", "Changed to Aria"); the one In use plays.

Compare plays two takes, A and B, at the same loudness on a loop. T flips between them at the same moment and Use B keeps B. A take no sound uses is deleted after 7 days; deleting one yourself is for good.

Edit a line by its wordsCopy Anchor Link

Words (in Edit) shows every word of a voice-over or conversation with its time. For an upload, Get the words writes them down first, priced.

• Click a word to play from it. While playing, the current word is bold.
• Delete words: select them and press Delete. Their audio is cut at the word edges, free and at once; undo brings them back.
• Change words by typing: Remake makes the sentence holding the change again, with the lines around it for an even flow, and splices it in at the word edges. Then Keep, Try again or Go back.
• Say it differently for a word said wrongly: spell it as it sounds, write its sounds in IPA between slashes (/ˈæləs/, Eleven v4), or put a tag before it.

Audio tagsCopy Anchor Link

Eleven v4 and v3 read tags in square brackets: [whispers], [laughs], [sarcastic], sounds like [applause]. Type [ (or 【) for the list, grouped as Feelings, Delivery, Reactions and Sounds; Pause adds …. Capitals stress a word. Multilingual v2 and Flash ignore tags: they're left out of what is sent and kept in your words.

VoicesCopy Anchor Link

Voices holds every voice you can use, each with a free preview.

A library voice can be withdrawn by its owner. Your takes stay; making more needs another voice.

Design a voiceCopy Anchor Link

Describe the voice ("A calm woman in her forties, light Scottish accent, warm and slow…"), give it words to say (100 to 1,000 characters) or press Write them for me, then Make 3 previews: charged once for all 3. Play them, pick one, name it and Keep this voice, free. Keep one before you close the dialog. More settings: the model (Voice Design v3 or v2), loudness, guidance, quality, seed, Improve my description, and on v3 a reference recording with its strength.

Adjust a voiceCopy Anchor Link

Adjust this voice on one of your voices, or on a library voice that allows it: describe the change ("Warmer and a little slower…"), pick a strength (Low, Medium, High or Max), then previews and Keep, as for a design.

Clone a voiceCopy Anchor Link

Three steps, consent first, before any recording is uploaded:

1. Whose voice is it? "It's my own voice", or "I have the speaker's permission" with the speaker's name, then I confirm this under the statement.
2. Your recordings: upload or record. 1 to 2 minutes of clear speech is best, at most 3 minutes, in up to 25 files and 100 MB. A mic check says when it's too quiet, too loud or noisy, and background noise is removed unless you turn it off.
3. Name it, with its language, accent, gender and age if you like, and Clone voice: free.

What consent means. You confirm the voice is your own, or that its speaker agreed to its cloning. Cloning anyone else is not allowed: their voice and every take made with it can be taken down. Our server checks it, not the page: a clone takes only the recordings a consent of yours names, given in the last 30 minutes. bInference keeps a record of the agreement (see Privacy). The director clones only from a recording you give it, after the same step, and won't imitate a real person by name: it offers a designed voice with the qualities you want instead.

Voice limitsCopy Anchor Link

• Up to 10 designed or adjusted voices and 2 clones at once.
• 6 voice saves (Keep or Clone) in 30 days, deleted voices included.
• A voice of yours unused for 30 days is removed. From 7 days before, Voices and every project using it say so, and using it once starts the 30 days again. Takes made with it stay; making more needs another voice.
• When the site's room for new voices is full, saving waits a while. Library and starter voices still work.

MixCopy Anchor Link

The timeline has tracks for Voice, Dialogue, Music, Effects and Atmosphere, and you can add more.

• Tracks: mute, solo, volume and pan.
• Sounds: move, trim, split and loop; gain from -24 to +12 dB; fades in and out (Linear, Smooth, Fast in or Slow in). Sounds that overlap on one track crossfade, or switch to Cut.
• Dip under voice: Music (on from the start) and Atmosphere step down while someone speaks: -6 to -24 dB (-15 to start), Fast, Normal or Slow.
• Loudness: the mix is set to one target when you export, every preset with true peaks at -1 dBTP. The preview plays the same mix, and the loudness chip shows what's playing and the export's target.

Undo works on every change, and versions keep each save point and each director run.

ExportCopy Anchor Link

Export mixes the saved version on our server at the loudness you pick.

• Each track as its own file (ZIP): every track as 24-bit FLAC with the master's level, so the stems add up to the mix.
• Subtitles: SRT and VTT from the words' times, at most 42 characters a line and 2 lines a cue. A sound without word times needs word timing first, priced.
• The AI label: every file carries Made with AI on bInference in its tags, with an export id that leads back to its account for a takedown. A song keeps its Content Credentials as a take; a mixed file is made anew, so its tags carry the label.
• Sounds still being made are left out, and named. You can close the dialog: the export carries on.
• Exports are free: one at a time, 30 a day and 300 minutes a day. A project exports up to 60 minutes. Files are kept 7 days.

Prices and limitsCopy Anchor Link

Prices are exact and shown before anything is made. Where a price says "up to" or "at most", it's the most you pay, and you're charged for what was made. A failed sound is free, and so is one cancelled while it waits.

PrivacyCopy Anchor Link

• Sent to the voice provider, only for the call that needs it: the words to say (with their tags), the prompts, lyrics and styles of songs and effects, voice descriptions, and the recordings you change, clean up, clone or have written down. The provider keeps a history of the requests it serves, under its own terms.
• The AI listening review hears a light, low-quality copy of the mix, sent like any AI call.
• Kept by bInference, private to your account: your projects, takes, the recordings a project uses, your voices and your exports. Projects stay until you delete them, then 30 days in Recently deleted; exports are kept 7 days. A recording you upload and don't use is deleted after a day. The request behind each sound, with its words, is deleted after 90 days. Our logs never hold your words, lyrics, prompts or file names.
• Consent records: for each clone, when you agreed, the statement's version and a fingerprint of its exact words, whose voice it is (and the speaker's name), fingerprints of the recordings, and your IP address and browser as keyed fingerprints. Never the recordings themselves. Kept as long as the voice and 1 year after it's deleted, for takedown requests. A consent no voice was made from is deleted after 30 days.

See the privacy policy.