Real-time speech-to-text in the terminal, using Soniox. Captions show up while people are still talking, tagged by speaker, and every session is saved as a transcript you can export later.
- Node.js 24 or newer
- A Soniox API key from the Soniox Console. Soniox projects belong to a region (US, EU, or Japan), and earpop has to be set to the same one.
- For live transcription: macOS (Apple Silicon or Intel), Windows x64, or Linux x64/arm64 with ALSA. On other platforms the transcript and API key commands still work.
npm i -g earpop-cliTo try it without installing, run npx earpop-cli.
npm only installs the microphone helper for your platform (one of the earpop-capture-* packages). Before each session, earpop checks the helper's SHA-256 hash against the one recorded at release.
Run earpop. The first time, it asks for your Soniox region and API key, verifies the key with Soniox, then asks you to pick a microphone and the languages people will speak. After that it goes straight to captions. ESC goes back from the key screen and quits from the others.
The language list starts from your system language plus English (English only on an English system). Press s and choose "Choose languages" to change them; type to search the list. Pick one language to keep the output in that language, or several for a mixed conversation.
To use other languages for a single run without changing your settings, pass --lang:
earpop --lang ja,en
earpop languages # list supported codes[S1] So the plan is to ship the beta on Friday.
[S2] Does that include the Windows build?
[S1] Yes, the capture helper is ready for x64.
● Listening en 03:12
p pause 1-9 speaker 0 all Up history c copy path m mic s settings ? keys ESC end
Captions fill the terminal, so a taller window shows more of the conversation. Dimmed text is provisional and gets replaced once Soniox finalizes it. A new line starts when the speaker or language changes, at the end of a sentence, or after a pause of more than two seconds. With more than one language selected, each line also shows the detected language, such as [S1 ja]. The status line shows the languages in use and any translation. Press Up or scroll the mouse wheel to read earlier captions from this session. Everything also goes into the transcript file.
| Key | Action |
|---|---|
p |
Pause or resume |
1 to 9 |
Show only that speaker. Press more numbers to add speakers, or the same number again to remove one. |
0 |
Show everyone again |
l |
With several languages or translation on, show only one language. Press again for the next; 0 shows all |
c |
Copy the transcript path to the clipboard |
m |
Switch microphone |
s |
Settings |
? |
Show every key |
Up |
Open the session history. Up/Down and PgUp/PgDn scroll, Home jumps to the start |
End |
Leave the history and return to live captions (ESC, or Down at the bottom, does the same) |
ESC |
Press twice to end the session, then once more to quit |
Ctrl+C does the same as ESC. When you quit, earpop waits for the last captions to be saved; press once more to quit right away. Keys also accept Korean jamo typed in Hangul mode (ㅔ for p, for example); input methods that hold a letter until the next key pass it on late, so switch to English for instant response. Captions keep running above the microphone and settings menus. On narrow terminals the key line keeps the most important keys; ? lists them all.
When you quit, earpop prints the transcript path and copies it to the clipboard.
earpop settings, or s during a session, lets you change the microphone, languages, translation, region, and API key. Microphone, language, and translation changes apply right away. Region and key changes apply the next time you run earpop. Each region has its own saved key.
earpop auth clear deletes the saved key for the current region.
Soniox can translate while it transcribes. In settings, choose "Translation" and pick one language to translate everything into it, or two to translate between them (for example English and Spanish in a bilingual meeting). Each translated line appears in italics right under the speech it translates, tagged with its language. For a single run:
earpop --translate ko # everything into Korean
earpop --translate en,es # English to Spanish and Spanish to EnglishTranslations are saved in the transcript too. Text exports label them, such as S1 (ko translation), and SRT and WebVTT exports show each translation at the same time as the speech, so players stack the two lines.
Names and jargon that Soniox should spell your way, such as product names or people, go in a term list. Write each term in the language and script people actually say it in. The list is sent with every session, and changes apply the next time you run earpop.
earpop terms add "Kold Brew" 웬민 Kubernetes
earpop terms # list
earpop terms remove Kubernetes
earpop terms clearEach session is a JSONL file named after its start time, such as 2026-10-12_14-03-27.jsonl. Final text is appended as it arrives, with timestamps, speaker, and language for every token.
earpop transcripts list # newest first, with start time and length
earpop transcripts view <id> # print as plain text
earpop transcripts export <id> # write <id>.txt next to the JSONL file
earpop transcripts export latest -f srt # newest transcript as subtitles
earpop transcripts export <id> -f md -o notes.md--format (-f) takes txt, srt, vtt, or md; --out (-o) picks the output path. Use latest in place of an ID for the newest transcript. Subtitle cues end at sentence boundaries and stay under about two lines.
The plain text version starts a new paragraph when the speaker or language changes, or after more than two seconds without speech. When a transcript mixes languages, each paragraph is labeled with its language code:
earpop transcript v1
id: 2026-10-12_14-03-27
started: 2026-10-12T14:03:27+09:00
model: stt-rt-v6
[00:00] S1: So the plan is to ship the beta on Friday.
[00:04] S2: Does that include the Windows build?
| Variable | Description |
|---|---|
SONIOX_API_KEY |
Used instead of the saved key |
SONIOX_REGION |
us, eu, or jp. Used instead of the saved region. Defaults to us. |
EARPOP_CAPTURE_BIN |
Path to a capture helper you built yourself. Skips the hash check. |
EARPOP_SKIP_CAPTURE_INTEGRITY |
Set to 1 to skip the hash check. Only for local builds. |
| macOS | Windows and Linux | |
|---|---|---|
| API key | Keychain, service com.earpop.app, account soniox-api-key-<region> |
~/.earpop/credentials-<region> |
| Settings | ~/Library/Application Support/com.earpop.app/cli-settings.json |
~/.earpop/cli-settings.json |
| Transcripts | ~/Library/Application Support/com.earpop.app/transcripts/ |
~/.earpop/transcripts/ |
Audio goes to the Soniox endpoint for your region and is never written to disk.
If Soniox rejects your key, the region is usually wrong. Run earpop settings, change the region, and enter the key again.
If a microphone stops sending audio, earpop reopens it and shows "No audio from the microphone; reopened it". Bluetooth headsets and hearing aids do this once when they switch to their microphone mode. After three tries in a minute it stops streaming and asks you to pick another microphone with m.
If you see "Capture helper integrity check failed", the helper doesn't match the released build. Reinstall earpop-cli.
On Linux, live transcription needs ALSA (libasound2) and access to the audio device. "System default" follows the input chosen in your sound settings, which is the right pick for most setups. Copying to the clipboard uses wl-copy, xclip, or xsel; without any of them earpop falls back to OSC 52, which not every terminal supports.
See AGENTS.md, docs/CONVENTIONS.md, docs/CONTRIBUTING.md, and docs/ARCHITECTURE.md.
MIT © wonmin