say Command on Mac: Wrong Voice Names and .mp3 Still Exit 0

The say man page on this Mac still shows say -v Alex -o hi -f hello_world.txt as its second example. I ran the voice part of it on macOS 26.4.1. Alex isn't installed here, and say didn't complain. It wrote a 43,796-byte file, returned exit status 0, and the audio was Yuna, the Korean voice. I then tried 14 more names that aren't in this Mac's voice list, including Siri, Ava, Zoe and a made-up "Nonexistent Voice 123". All 14 did the same thing: exit 0, same bytes, Yuna.

That is the pattern for most of what goes wrong with the say command on Mac. The man page says say "returns 0 if the text was spoken successfully, otherwise non-zero", and in my tests a lot of failures still came back as 0: a .mp3 output that is 16 bytes long, an .ogg that is empty, a muted speaker, an audio device ID that doesn't exist. Below is what I measured on a Mac mini with /usr/bin/say (197,824 bytes, man page dated 2020-08-13): every voice it lists, 24 option and error cases, eight speech rates on four voices, the volume and pause commands on all 72 voices, and a count of the 58 Stack Exchange questions people have asked about say.

What the say command does

say converts text to speech with the system synthesizer. It speaks through the current output device by default, or writes an audio file with -o. Text comes from the arguments, from a file with -f, or from standard input when you give it neither, so echo hello | say works. The options that matter day to day are -v for the voice, -r for the rate in words per minute, -o for an output file, -a for an audio device and -i to highlight words as they're spoken. The full list is in the say(1) man page.

Two things trip people before any of that. First, text that starts with a dash is read as an option: say "-5 degrees outside" exits 1 with invalid option -- 5. Put -- before the text and it's spoken. Second, the man page tells you to list things with a bare question mark, as in say --file-format=?. In zsh, the default shell since Catalina, that line fails before say runs: zsh: no matches found: --file-format=?. Three of the four ? examples in the man page fail this way in zsh. Quote it: say -v '?', say '--file-format=?'. Bash passes an unmatched ? through, so it works there until the folder has a file with a one-character name. I put a file called a in an empty folder and ran say -v ? in bash: the glob turned it into say -v a, which printed no list, spoke in a fallback voice and exited 0.

Voices: 72 listed, and the ones that are missing

say -v '?' lists 72 voices on this Mac, across 51 locales. 27 are English, and 20 of those are en_US. That list is not the same as what apps see. AVSpeechSynthesisVoice, the API most apps use, returns 68. The four that only say lists are Aman, Tara, Ona and Aru, and three of them are backed by "neuralAX premium" voice assets under /System/Library/AssetsV2. The voice assets that are installed but unusable are bigger: Minji (Korean, two Siri assets of 218 MB and 55 MB) and Simone (en_US, 85 MB) are on disk, and say -v Minji and say -v Simone both fall back to Yuna.

That matches the long-running Stack Overflow question Make the say terminal utility work with Siri voices: Siri voices can't be picked with -v, but since Ventura say uses one if it's set as the system voice and you leave -v off. My default output, with no -v, matched none of the 72 listed voices byte for byte, which is consistent with that. I didn't confirm by ear which voice it was.

What I passed to -v	Exit	What it spoke
Samantha, samantha	0	Samantha (names are case-insensitive)
com.apple.voice.compact.en-US.Samantha	0	Samantha (full identifiers work)
Alex, Victoria, Agnes, Bruce, Vicki, Tom, Susan, Allison	0	Yuna (not installed here)
Siri, Ava, Ava (Premium), Zoe, Eddy	0	Yuna
Minji, Simone (Siri assets installed)	0	Yuna
Deranged, Hysterical, Princess	0	Yuna
Nonexistent Voice 123	0	Yuna

The Deranged row is a rename trap. AVSpeech still reports the identifiers com.apple.speech.synthesis.voice.Deranged, .Hysterical and .Princess, but the voices are now listed as Wobble, Jester and Superstar, and old scripts that use the old names get the fallback voice. Why Yuna? This account's only system language is Korean, and Yuna is the ko_KR voice, so the fallback looks like "the default voice for your language". On an English Mac you'd probably get an English voice, which makes the mistake harder to hear. One caveat: in my very first batch, Alex, Siri and Eddy produced a 41,498-byte file that matched no listed voice. Every run after that, 33 more, was Yuna.

Downloading more voices happens in System Settings, not the terminal: Accessibility, Spoken Content, System voice, Manage Voices. Apple's page for that panel is Have your Mac speak text that's on the screen. Once a voice is downloaded, check that say -v '?' lists it before a script relies on it.

Output files: which extensions actually work

say picks the file format from the extension you give -o. say '--file-format=?' lists 18 formats here, but a listed format isn't a promise. I wrote "hello world" with each extension and opened every file with afinfo.

Command	Exit	Result
-o hi.aiff or -o hi	0	42,950 B AIFF; no extension gets .aiff appended
-o hi.m4a	0	42,950 B, uncompressed PCM inside an M4A box
-o hi.m4a --data-format=aac	0	AAC; 1.35 MB vs 14.2 MB AIFF on a 756-word text
-o hi.aac	0	4,039 B ADTS AAC
-o hi.flac	0	42,705 B FLAC (barely smaller for speech)
-o hi.wav	1	Opening output file failed: fmt?
-o hi.wav --data-format=LEI16@22050	0	42,950 B WAV
-o hi.mp3	0	16 bytes, no audio
-o hi.ogg, .opus, even with --data-format=opus	0	0 bytes
-o hi.txt	0	A 42,950 B AIFF named .txt

Two of those surprise people most. The plain .m4a is the same size as the AIFF because say only compresses when you ask with --data-format=aac. And there is no MP3 path at all; the "say command mp3" answers all pipe the AIFF through lame or ffmpeg. If you want a small file without extra tools, use .m4a --data-format=aac, which was 10.4 to 10.7 times smaller than AIFF with the four voices I tried on a 756-word text (the say man page itself).

Rendering to a file is fast for most voices. The 756-word text became 323 seconds of audio with Samantha in 1.87 seconds of wall time, about 173 times faster than real time. Daniel was 192 times, Fred 98. Aman, one of the neural voices, took 21.0 seconds for 343 seconds of audio: only 16 times real time. If you batch long files with a neural voice, plan for that.

Speed: what -r actually does

-r is documented as words per minute, but it behaves like a setting with limits that depend on the voice. Leaving -r off gave exactly the same length as -r 175 for all four voices I tested, so 175 is the default. The same 18-word sentence came out at 5.0 seconds with Samantha, which works out to 216 words per minute including the silence at both ends, not 175.

At the slow end, -r 1 doesn't mean one word per minute. Samantha took 7.5 seconds, only 1.5 times slower than the default. -r 90 and -r 100 came out at the same length for Samantha and Daniel. At the fast end, Samantha, Daniel and Aman all stopped at 1000: -r 5000 gave the same length as -r 1000. Fred kept going, to 0.65 seconds. Invalid values aren't rejected: -r 0 and -r abc were silently treated as the default, exit 0. -r -50 made speech slower. For a slow change inside a sentence, the embedded command [[rate 400]] worked on all four test voices.

Volume and pauses

say has no volume flag. The documented way is an embedded command in the text: say "[[volm 0.3]] quieter now", where 1.0 is full volume. I rendered a sentence with and without [[volm 0.05]] on all 72 voices and compared the RMS level of the two files. 69 voices got quieter. Three didn't change at all: Aman, Tara and Ona gave byte-for-byte identical audio with and without the command. Those are the three neural voices from the asset list above, so a script that lowers volume this way will be loud on exactly the voices people pick because they sound best. The drop also isn't linear: [[volm 0.3]] brought Samantha to 14% of her normal RMS, not 30%.

Pauses worked everywhere. [[slnc 1000]] asks for 1,000 milliseconds of silence; across the 72 voices it added between 0.55 and 1.05 seconds. If you need exact gaps, render the pieces separately and join them.

say command not working: the causes I could reproduce

"say command not working mac" was the most frequent completion in my Google autocomplete collection for this command. Here are the causes I could reproduce on this Mac. All of them exit 0 except a bad device name and the .wav case.

• Output is muted. osascript -e 'get volume settings' on this headless Mac mini returned output muted:true. say "test" still ran for 1.15 seconds and exited 0. Nothing in say's output tells you the speaker is muted.
• The voice isn't installed. You hear a different voice, or one in a different language, as in the table above.
• The text ended up in the wrong place. say -o "Hello, this is not working" treats the quoted text as the file name and reads speech from standard input. From a script with no input it wrote Hello, this is not working.aiff, 4,096 bytes with zero seconds of audio. In a terminal it waits for you to type, which looks like a hang. That is the accepted answer on MacOS is not saving output from say -o.
• A bad device ID. say -a 9999 "test" exited 0. A bad device name is caught: Found no Audio Output Device matching `No Such Device', exit 1. List devices with say -a '?'.
• The file format. .mp3, .ogg and .opus give you a file with no audio and exit 0. .wav without --data-format is the one that fails loudly.

The fix in the old answers to The say Command is not Working in High Sierra was kill `pgrep speechsynthesisd`. That process doesn't exist on macOS 26.4.1. While say was in use, the speech processes running here were SpeechSynthesisServerXPC (two instances), WardaSynthesizer and sirittsd. I didn't have a hang to test, so I can't say which one to restart if say freezes.

A say wrapper that checks what say doesn't

This is the function I'd put in a script that has to produce a usable file. It refuses unknown voices instead of falling back, passes -- so text can start with a dash, compresses to AAC, and checks that the file has audio in it.

# say with the checks it doesn't do itself
speak_to_file() {  # usage: speak_to_file VOICE OUT.m4a TEXT...
  local voice="$1" out="$2"; shift 2
  if ! say -v '?' | sed -E 's/ +[a-z]{2,3}_[A-Z0-9]+ +#.*$//' | grep -qxF -- "$voice"; then
    echo "speak_to_file: voice not installed: $voice" >&2; return 2
  fi
  say -v "$voice" -o "$out" --data-format=aac -- "$@" || return
  local secs
  secs=$(afinfo "$out" | awk '/estimated duration/ {print $3}')
  if ! awk -v s="${secs:-0}" 'BEGIN { exit !(s > 0) }'; then
    echo "speak_to_file: $out has no audio" >&2; return 3
  fi
}

I tested it in bash and zsh. speak_to_file Alex out.m4a hello returns 2 here. Names with spaces like "Bad News" work because the sed strips the locale column instead of splitting on spaces. An empty string returns 3, since say writes a 4,096-byte file with no audio for that. It doesn't catch a muted speaker, because it never plays anything. For live speech from a job that runs while you're away, keep the Mac awake while it talks; caffeinate on Mac covers the timeout flag that catches people. To speak the clipboard, pbpaste | say works, with the encoding caveats in pbcopy on Mac.

What 58 Stack Exchange questions ask about say

I pulled every question on Ask Different, Super User and Stack Overflow whose title mentions say as a command, through the Stack Exchange API, and removed the ones about Twilio's Say verb and the English word. 58 questions were left, from January 2010 to October 2025, with 330,754 views combined. I sorted them by what the asker wanted.

Topic	Questions	Views	Share of views
say on Linux or Windows, Python, licensing	6	113,661	34%
Voices, languages, Siri and Personal Voice	11	74,810	23%
Not working: cron, at, logged out, AirPlay, upgrades	9	46,820	14%
Output files, formats, MP3	11	38,881	12%
Scripting and pipes	11	23,062	7%
Rate and volume	4	17,885	5%
Pronunciation	6	15,635	5%

The biggest group is people who want say somewhere else; the top question, an Ubuntu equivalent, has 62,868 views on its own. After that, voices lead, which fits the fallback behavior above: a wrong voice doesn't produce an error to search for, so people ask why say "sounds wrong". A recent one, Why does say think my Personal Voice is in Spanish?, describes a Personal Voice listed as es_MX that speaks English when the screen is unlocked and Spanish when it's locked. I didn't have a Personal Voice to test. Quietly doing the wrong thing isn't unique to say: in plutil on Mac, -bool 1 saves false. hdiutil on Mac is the opposite case, where a documented refusal arrives as a crash.

FAQ

How do I change the voice of the say command on Mac?

Use say -v NAME "text", for example say -v Samantha "hello". Run say -v '?', with the quotes in zsh, to list installed voices. If the name isn't installed, say doesn't report an error; it speaks in a fallback voice and exits 0. Add voices in System Settings, Accessibility, Spoken Content, System voice, Manage Voices. Siri voices can't be chosen with -v, but say uses one when it's set as the system voice and you leave -v off.

How do I save say output as an MP3 or audio file?

say can't write MP3. A .mp3 output name gives a 16-byte file with no audio and exit 0. Write AAC instead with say -o out.m4a --data-format=aac "text", or AIFF with say -o out.aiff "text". For WAV you must add a data format, such as --data-format=LEI16@22050. Convert to MP3 afterward with ffmpeg or lame if you need it.

How do I change the speed or volume of say?

Speed is -r in words per minute; the default is 175. Many voices stop getting faster around 1000, and 0 or non-numbers are ignored. There is no volume flag. Put [[volm 0.5]] at the start of the text, where 1.0 is full volume. That works for most voices, but on macOS 26 the neural voices Aman, Tara and Ona ignore it.

Method: all tests ran on 2026-10-08 between 16:30 and 17:00 KST on a Mac mini M4 (Mac16,10, macOS 26.4.1 build 25E253, /usr/bin/say 197,824 bytes), in a scratch folder under my work directory, from a launchd-started job in the logged-in session. The account's only system language is Korean. Voices were compared by MD5 of the rendered AIFF for "hello world"; durations come from afinfo; volume from the RMS of each file after afconvert to 16-bit WAV. Render speeds are single wall-clock runs. The AVSpeech count comes from a small Swift program calling AVSpeechSynthesisVoice.speechVoices(), and the asset sizes from du on /System/Library/AssetsV2. Live playback ran three times, all on a muted output, and I didn't judge any audio by ear; everything else was rendered to files. Not tested: Personal Voice, SSH sessions, cron, AirPlay, and English-language Macs, where the fallback voice may differ. The 58-question count comes from the Stack Exchange API title search on the three named sites on 2026-10-08, filtered and grouped by hand from the titles. Search phrasings like "say command not working mac" come from my Google autocomplete collection for five say queries on the same day.