Who said what, when.

Drop in an interview, a podcast or a call. Murmur writes down every word, tells the voices apart and lays them out on a timeline you can play, mark and cut. Not a second of audio leaves your Mac.

Download for Mac Try it in your browser

Version 0.1.5 · macOS 26 or later · Apple silicon · Free

4 MB app. The speech models, 515 MB, download once on first launch.

No Mac? Murmur in the browser transcribes and finds the speakers in Chrome, Edge or Safari, without uploading the file. It's slower, records only the microphone and browser tabs, and leaves out word re-timing and summaries.

Murmur's project window. A waveform colored by speaker runs under three named marks, with one lane each for Maya, Jonas and Producer. Next to the play controls is Enhance; the transcript header holds the language, Fillers and Redact. The transcript has the word being played highlighted, and a sidebar shows the summary and each speaker's talk time.

Every word has a time and a voice.

Below are forty seconds of a recording with synthetic voices, exactly as Murmur transcribed them. Click a word to hear it. Marking works as in the app: once where a part starts, once where it ends, and it’s ready to copy.

00:00.0 / 00:40.0
Space Play or pause ←→ Five seconds M Start or end a mark Esc Cancel

Marks

    
        
    
      
    
    
    
  
    

  
    
    
    
    
      
    
        
    A small window for the call you’re in.
    
        
    Start a recording and Murmur keeps two tracks: your microphone as Me, and whatever your call app plays as Others. The transcript runs live in a compact window that stays on top.
    
        
    
          
    Invisible when you share your screen.
    The window doesn’t show up in the shared picture. Settings can change that.
    
          
    ⌥⌘M from any app.
    Marks the moment without leaving the call.
    
          
    Speakers when you stop.
    The other voices are told apart and the recording opens as a project, timeline and all.
    
        
    
      
    
      
    
        
          
          The compact recording window at 23:41: level meters for Me and Others, the live transcript with two marked lines, and Mark, Pause and a red Stop button. A note at the bottom says it’s hidden from screen sharing.
        
      
    
      
    
        
          
          The recorder shrunk to a strip: a timer, the latest line of the transcript, a mark button and Stop.
        
        
    Or shrink it to a strip: the timer, the latest line, Stop.
    
      
    
    
    
  
    

  
    
    
    
    
      
    Ten minutes of audio. About ten seconds.
    
      
    Murmur doesn’t send your recording to a big model somewhere. It runs six small ones on your Mac, one after another. No account, no upload, no subscription.
    
      
    
        
      
          
    1. Decode0.6 s
      2. 
          
    2. Language Ear1.5 s
      3. 
          
    3. Words Voz1.7 s
      4. 
          
    4. Speakers Who1.6 s
      5. 
          
    5. Timing Align1.5 s
      6. 
          
    6. Fillers Uhm3 s
      7. 
        
    
        
    Ten minutes of audio on an M2, Release build. Tongue then checks the language against the words.
    
      
    
      
    Speech models by Desert Ant Labs. Voz is based on NVIDIA Parakeet TDT 0.6B v3. Loading a model sends an anonymous usage count to Desert Ant Labs for licensing. Your audio is never sent.
    
    
    
  
    

  
    
    
    
    
      
      
    
        
    Fix words in place
    Type over a misheard word and Murmur re-times it against the audio, so highlights and cuts stay exact. Rename, merge and reassign speakers. Everything can be undone.
    
        
    Export
    Plain text, text with timestamps, Markdown, SRT, WebVTT, word-level JSON. Or audio and video clips: one file per mark, or all of them joined.
    
        
    Hide personal details
    Names, phone numbers, emails, addresses and IDs, hidden in the transcript, copies and exports. Exported audio clips get a tone over them.
    
        
    25 languages
    Recognized from the sound, then checked against the words. For any other language Murmur offers Apple’s on-device transcriber.
    
        
    Fillers and noise
    Um and uh are found and left out of exports. Optional enhancement cleans up a noisy room for listening, transcribing and export.
    
        
    Titles and summaries
    Where Apple Intelligence is available, it names each project and writes a short summary, which exports can include.
    
        
    Projects, not copies
    A .murmur file points to your original audio or video instead of copying it, and finds it again if you move it. Move a project to the Trash from the library and a recording’s audio goes with it; imported files stay where they are.
    
        
    Colors you can tell apart
    Nine speaker color schemes, including a color-blind safe one, each tested for contrast in light and dark.
    
      
    
    
    
  
    

  
    
    
    
    
      
    
        
        
    
          
    Murmur 0.1.5
    
          
    Free, for macOS 26 or later on Apple silicon. Notarized by Apple, and it updates itself.
    
          
    
            
              
              Download for Mac
            
            Changelog
            Try it in your browser
          
    
          
    
        
    
      
    
      
    
        
    Or build it
    
        
    brew install xcodegen
git clone https://github.com/djui/murmur
cd murmur && make run