Voice · record and perfectIn developmentWindows and Linux app first

Sound Like You.
On Your Best Day.

Voice records anything with a voice in it, from song vocals to audiobooks and podcasts. It cleans the recording, levels it, tunes it and exports it for where it's going. It repairs your recording and never regenerates your voice, and all of that processing runs on your device. Every demo below runs in your browser on generated audio, and makes a sound only when you press play. The screens are from Voice's browser test build, and every number says how it was measured.

  • Windows
  • Linux
  • macOS
  • iOS
  • Android
All planned. Downloads open through the waitlist.
Join the Voice waitlist What's built, what's not
01 · Clean

Clean the room, keep the voice.

One control, Off to Strong. Clean cuts the rumble, finds the mains hum and notches it out, repairs clicks, and takes the room's noise down against a noise print it learns from your quietest moments. The noise goes down to a floor, never to dead silence, so the voice keeps its air. Breaths stay unless you choose otherwise. Every step reports what it did, and each can be undone on its own. Try it on a generated phrase with noise, hum and two clicks in it.

Grey is as recorded; colour is after Clean. Marks show the clicks it repaired. A simple version of Voice's chain, running in your browser.

Clean
Listen to
What Clean did

    Before and After are level-matched, so louder never sounds like better.

    Rumble and hum

    Built

    A high-pass at 80 Hz by default, and narrow notches on 50 or 60 Hz and its harmonics, found by listening to the quietest seconds.

    Clicks and sibilance

    Built

    Short spikes far above their surroundings are drawn through smoothly; “s” sounds are softened between 4.5 and 16 kHz only when they dominate.

    Breaths

    Built

    Found and kept by default. Soften or remove them on request; laughs and breaths are part of a voice.

    Repair, never regenerate

    Built

    Clean is signal processing on your device, so a cleaned recording is still yours: its AI declaration stays None.

    Voice's Edit screen in Easy mode: a 10-second recording's waveform, Before and After, and four controls, Clean at Medium, Level, Warmth and Tune. Under them: Noise down 17 dB, hum 60 Hz out, 2 clicks fixed. A bar says it exports for podcast at −16 LUFS, now −13.9 LUFS.
    Clean at Medium in Voice's browser test build, on the same kind of generated phrase. The findings line is what Voice measured on it.
    02 · Tune

    In tune, still human.

    Tune finds each sung note and its key, then moves the note's centre to the scale and keeps how you sang it: the vibrato, the drift, the slides. Amount says how far it moves; Speed goes from natural to robotic. In Pro you can move a single note or straighten its drift. Tune is sound processing, so it doesn't raise the AI declaration. Drag Amount and Speed, then listen.

    Seven sung notes on a piano roll in A minor, each a little sharp or flat, with their vibrato.

    The lines are the pitch as sung (grey) and after Tune (colour). Notes and their cents off the scale are listed beside.

    Tune · A minor (found) Listen to
      Within 2 centswhere notes sung 20 to 40 cents off land after Tune, with the vibrato kept
      No clicksat note edges after tuning
      Nonethe AI declaration after Tune and Clean: both are sound processing (decided 22 October 2026)

      Measured 22 October 2026 on test notes sung 20 to 40 cents off, through Tune on the range finder's pitch tracker (YIN), in the audio core that the browser build and the desktop app share. Live Tune, which tunes while you record, isn't built yet.

      Voice's Edit screen in Pro mode: the waveform with an overview, and under it Notes: 7 notes, A minor (found), each on a piano roll with its wavy pitch line, A3, C4, D4, E4, D4 and C4. On the right, Cleanup with Clean, Noise, Clicks, Sibilance, Rumble cut, Hum and Breaths, Level and tone, and Tune with Amount 70%, Speed Gentle and Key Find it (A minor).
      Tune in Pro, from Voice's browser test build, on a generated phrase sung 22 to 38 cents off. It found the key and all seven notes.
      03 · Tone

      Warmth you can see and hear.

      Warmth adds a little body around 200 Hz and a softer top, and Rumble cut takes out what sits under a voice. In Pro every setting has a number. Drag the points on the curve, or use the sliders, while the phrase plays.

      The tone curve from 20 Hz to 20 kHz, with a rumble cut, a warmth shelf and a top shelf.

      Drag a point. The curve is drawn from the same filters you hear.

      Tone

      Easy and Pro

      Built

      Easy shows Clean, Level, Warmth and Tune with Before and After. Pro adds every setting with its number. Switching never changes the sound.

      Level

      Built

      A slow leveller brings quiet phrases up and loud ones down, holding through pauses so the room is never pumped up.

      Vocal EQ

      Planned

      Bands that follow the pitch, and matching to a reference track you own, analysed on the device.

      04 · Export

      Ready for where it's going.

      Pick where it's going: a podcast, an audiobook for Audible (ACX), a vocal for a song, or a voice note. Voice measures what the export will deliver, before anything is written, and every failing check comes with its fix. WAV, FLAC and MP3 are built; a .voice file keeps the takes and every edit in one file. Check the phrase against a preset.

      Generated phrase · 6 s · mono

        Where is it going?

        Measured here with the loudness method podcasts and streaming use (ITU-R BS.1770, gated), on sample peaks. Voice measures true peaks.

        3.4 sto pack 3 minutes of stereo into a .voice file, about a minute per hour, in the background
        26 msto draw a recording's waveform, at any length
        1 secondthe most a crash mid-take can lose: the take is synced to disk every second

        Packing and waveform measured 19 October 2026 on the build machine, a release build of the audio core. The crash limit is the recorder's design, tested by killing a real recording process and checking every surviving sample. cargo test in avlory/packages/avlory_audio_core

        Voice's Export dialog: Where is it going? Podcast at −16 LUFS with peaks under −1 dBTP, Audiobook (ACX), Music vocal, Voice note and As it is. Checks, before anything is written: Loudness passes, now −13.9 LUFS, exported −16.0 LUFS; True peak passes. File name, format WAV 24-bit, and the AI declaration, measured as None.
        The Export dialog, from Voice's browser test build: each check is measured on what the export delivers.
        05 · Transcripts

        Edit the words, the audio follows.

        Voice transcribes on your device, with every word timed. Select words and the audio follows; cut them and the cut gets short fades. It finds fillers and long pauses for you to take out, each one put back on its own. Export subtitles as SRT or WebVTT, or plain text. The speech model downloads once, after you say so, and is checked before each use. Take out the fillers in an example podcast intro.

        Episode 12 · intro

        
        
        
        
        
          Fillers and pauses
          
          
          
          
          
          An example transcript. Click any word to cut it or put it back.
          
        
          
      
          
      
          
        
          0.10×real time to transcribe with the standard model (Whisper tiny.en, 39 MB)
          
        
          0.20×with the more accurate model (base.en, 74 MB), a setting
          
        
          5.3%words wrong on LibriSpeech clips, with either model
          
      
          
      
          Measured 22 October 2026 in the browser test build, both models on the same clip. The model is OpenAI's Whisper (MIT licence) run in Avlory's own Rust code, downloaded from avlory.com and checked against its published SHA-256. Transcripts are analysis, so the AI declaration stays None.
          
    
          
  
          

  
          
    
          
      
          06 · Live sessions
          Record together, apart.
          Built, on a testing relay only
          Talk live from different places while each person is recorded on their own device at full quality, so the recording never depends on the network. Everyone agrees to being recorded before they join, signs in with an Avlory ID, and sees the recording light while it's on. Afterwards the takes line up exactly, even though every sound card's clock runs a little fast or slow. Set a guest's clock and line the takes up.
          
      
          
        
          
          
          A one-minute session · five claps, zoomed inNot lined up
          
          Five claps heard by the host and the guest, each window 300 ms wide.
          
          
        
          
        
          
          Guest's sound card
          
          
          
          
          
          Simulated, with Voice's method: sync markers every tenth of a second, then a straight-line fit gives where the take starts on the host's clock and how fast its card really ran.
          
        
          
      
          
      
          
        
          97 to 103 msmouth to ear, the live conversation as Opus at 48 kHz (target under 150 ms on one network)
          
        
          0.06 to 0.40 mshow closely the takes line up afterwards (the test allows 2 ms)
          
        
          32 kbit/sthe conversation stream, about a third of the data of the earlier ADPCM stream
          
      
          
      
          Measured 29 October 2026 on the build machine in 15 runs, in one process with stand-in devices and 6-second takes, the guest's clock running about 500 ppm fast. Over the internet the network adds 20 to 100 ms. Not yet measured between real computers.
          
      
          
        Diagram: each person records on their own device and uploads as they go; the live conversation is only for hearing each other.
          
            Host's devicerecords the host, full quality
            Guest's devicerecords the guest, full quality
            Relaysealed; can't read either
            
            talktalk
            
            The host's .voicetakes lined up, consent kept
          
        
        
          Diagram: how a live session is recorded. The takes upload in sealed chunks while you talk, and resume after a dropout.
          
      
          
    
          
  
          

  
          
    
          
      
          07 · Your voice
          Your voice stays yours.
          The range finder listens to you sing and tells you your lowest and highest notes, where you're comfortable, which voice type that falls into, and keys that suit you, so you can write around your voice. It's a guide, not a label. And the rules about voices are simple: yours is yours, and someone else's is used only from recordings they made. Try the range finder on three generated singers.
          
      
          
        
          
          
          Your voice
          
          
          
          
          
          A keyboard from E2 to C6 with the notes reached and the comfortable range.
          
          
        
          
        
          
          Singer
          
          
          
          
          
          Voice finds the notes by listening (YIN pitch tracking, on the device). Here the notes are the ones the demo generated; the ranking is Voice's, copied from its code.
          
        
          
      
          
      
          
        
          Doubles and harmonies
          Planned
          From your own voice, always fine.
          
        
          Someone else's voice
          Decided
          Only from recordings they made: a .voice file they sent you, or a session you recorded together.
          
        
          Cloning your own voice
          Planned
          Fix a flubbed word without re-recording. Only your own voice, trained and run on your device, and everything it makes is marked as generated.
          
        
          Stock narrators
          Planned
          Licensed synthetic voices, not cloned from real people, on the device and always marked as generated.
          
      
          
      
          
        Voice's Find my range screen: the note being sung, A3 at 224.3 Hz and +33 cents, a pitch trace over a piano keyboard, and the result: Tenor, leaning baritone. You reached A3 to E4, and you're comfortable from A3 to D4; 7 notes sung. A keyboard shows the notes reached, the comfortable range and time on each note, and a Write in my range button.
        
          Find my range, from Voice's browser test build, listening to a generated phrase through a stand-in microphone.
          
      
          
    
          
  
          

  

  
          
    
          
      
          Status
          Where Voice is today.
          In development. The desktop app for Windows and Linux records, cleans, levels, tunes, transcribes and exports, with chapters and markers, the range finder, Send to Sound and live sessions on a testing relay; a browser build is for testers. Live Tune, the player embedded in Quill, AAC and Opus export, the phone apps and own-voice cloning come later. Downloads open through the waitlist.
          
      
          
        
          Tested
          • On a Linux build machine with no sound card, using stand-in devices.
          • Opening a recording brings back takes a crash left and unsaved changes; the core recorder survives being killed mid-take.
          • Every control labelled, the whole loop by keyboard, 200% text, light and dark.
          
        
          Not tested yet
          • Real microphones, interfaces and latency on the test computers.
          • Killing the real app mid-take (the core recorder is tested).
          • Live sessions between real computers, and macOS, iOS and Android builds.
          
      
          
      
            
        
          • No feature moves up a tier
            A saved project always opens and exports, even after a plan ends.
            • 
        
          • Never locked out offline
            A failed licence check never stops editing or export.
            • 
        
          • AI never changes work silently
            Every AI change is a visible, undoable suggestion in the AI ledger.
            • 
        
          • Free to start
            Plans and prices are announced at launch. All the promises
            • 
      
          
    
          
  
          

  
          
    
          
      
          Waitlist
          Try Voice First.
          Voice's downloads open through the waitlist. We'll email you when it's ready on your platform.
          
      
          
        
          
          
          
            Which apps?
            
          
              
              
              
              
            
          
            
          
          
          
          
          
            
            
            
          
          
          
          
          
            
            
            
          It helps us decide what to finish first.0 / 500
          
          
          
          
          
          
          We'll use your email only to tell you when the apps you picked open for download. Read the privacy policy.