gemgen

Gemini-TTS CLI for the Google Cloud Text-to-Speech public browser demo. It has one command, gemgen tts, and always forces voice.modelName to gemini-3.1-flash-tts-preview.

deno run -A cli.ts tts --text "Hello" --out speech

Run or install from JSR after publishing:

deno x -A jsr:@cliat/gemgen/cli tts --text "Hello" --out speech
deno install -g -A -n gemgen jsr:@cliat/gemgen/cli

The command launches headed Google Chrome through playwright-cli on PATH every run, opens https://cloud.google.com/text-to-speech, uses the embedded demo context, waits for any CAPTCHA in the visible browser, decodes audioContent, writes the next sequenced file for --out, and closes the browser. Install the browser driver once if needed:

npm install -g @playwright/cli
playwright-cli install-browser --browser chrome

Commands

gemgen tts --text "Hello" --out speech
gemgen --json tts -t "Hello" -p "Read warmly." -v Achernar -l en-US -e LINEAR16 -o output
deno x -A jsr:@cliat/gemgen/cli --json tts -t "Hello" -o output
gemgen tts --json-template > request.json
gemgen tts -i request.json -o speech
deno run -A cli.ts tts --help

--json prints one stable success object to stdout. Progress, CAPTCHA instructions, and errors go to stderr. outputs[] lists every file written.

On Windows, run headed browser generation directly in a visible console window. Avoid wrapping the gemgen process in PowerShell pipes or Tee-Object; those can interfere with browser-launch handles.

Library import:

import { createTtsJsonTemplate, parseTtsJsonInput } from "jsr:@cliat/gemgen";

Options And JSON

Granular flags override JSON fields. JSON input overrides defaults. --out is CLI-only and is never read from JSON. input.text is an array in JSON; each string is submitted as a separate service call with the same settings. --text accepts one string and overrides JSON text. gemgen waits 5-10 seconds before each next service call and retries transient demo/proxy failures for the same item.

Flag	JSON field	Default	Notes
-t, --text <text>	input.text[]	none	CLI accepts one string. JSON accepts a non-empty string array. Cannot combine with structured turns.
-p, --prompt <text>	input.prompt	omitted	Style instructions.
-v, --voice <name>	voice.name	Achernar	Single-speaker Gemini voice.
-l, --language <code>	voice.languageCode	en-US	BCP-47 language code.
n/a	voice.modelName	forced	Always sent as gemini-3.1-flash-tts-preview. JSON values are ignored.
-e, --encoding <value>	audioConfig.audioEncoding	LINEAR16	LINEAR16, ALAW, MULAW, MP3, OGG_OPUS, PCM.
-r, --speaking-rate <number>	audioConfig.speakingRate	1	Range 0.25..2.0.
-P, --pitch <number>	audioConfig.pitch	0	Range -20..20.
-g, --volume-gain-db <number>	audioConfig.volumeGainDb	0	Range -96..16.
-s, --sample-rate <hz>	audioConfig.sampleRateHertz	omitted	Positive integer hertz.
--profile <id>	audioConfig.effectsProfileId[]	[]	Repeatable; applied in order.
--speaker <alias=voice>	voice.multiSpeakerVoiceConfig.speakerVoiceConfigs[]	[]	Repeatable; use only with structured turns. Alias must be alphanumeric.
--turn <alias:text>	input.multiSpeakerMarkup.turns[]	[]	Repeatable structured dialogue turn. JSON uses { "speaker": "...", "text": "..." }.
--turns-file <path>	n/a	omitted	JSON array of { "speaker": "...", "text": "..." }; replaces repeated --turn values.
--start-at <number>	n/a	1	Resume text-array input from the 1-based item number.
-i, --input <path>	full request object	omitted	Reads JSON shaped like --json-template.
--json-template	n/a	n/a	Prints a full JSON template/example and exits.
-o, --out <path>	n/a	required	Output stem. Creates parent dirs and writes the next numbered file.
--json	n/a	false	Stable JSON success output.

--profile maps to audioConfig.effectsProfileId. --speaker Sam=Kore maps alias Sam to Gemini voice Kore for structured dialogue. --turn Sam:Hello appends { "speaker": "Sam", "text": "Hello" }. --turns-file turns.json reads the same array shape used by input.multiSpeakerMarkup.turns.

--out path/to/file scans for path/to/fileNNNN.<ext>, creates the parent directory if needed, and writes the next number. If path/to/file0004.wav exists, -e LINEAR16 --out path/to/file writes path/to/file0005.wav. Known audio extensions on --out are stripped, so --out speech.wav still uses the stem speech.

Each text item is checked before the browser opens: input.text[] items must be at most 4,000 UTF-8 bytes, input.prompt at most 4,000 UTF-8 bytes, and text plus prompt at most 8,000 UTF-8 bytes.

If a batch stops after writing some files, resume from the next item:

gemgen tts -i request.json -o speech --start-at 4

JSON Template

gemgen tts --json-template > request.json
gemgen tts -i request.json -o speech

Template JSON has no output path. Defaults target falling-asleep videos: uncompressed LINEAR16, 48 kHz, neutral audio with no device profile, and a calm soothing narration prompt.

{
  "input": {
    "text": [
      "Paste the first narration segment here.",
      "Paste the next narration segment here."
    ],
    "prompt": "Calm, soothing narration. Slow gentle pacing, soft warmth, relaxed clarity, and peaceful pauses."
  },
  "voice": {
    "languageCode": "en-US",
    "name": "Umbriel",
    "modelName": "gemini-3.1-flash-tts-preview"
  },
  "audioConfig": {
    "audioEncoding": "LINEAR16",
    "speakingRate": 1,
    "pitch": 0,
    "volumeGainDb": 0,
    "sampleRateHertz": 48000
  }
}

--json output:

{
  "ok": true,
  "command": "tts",
  "modelName": "gemini-3.1-flash-tts-preview",
  "outputs": [
    { "out": "speech0001.wav", "bytes": 12345, "index": 1 }
  ]
}

Examples

gemgen tts -t "Welcome aboard." -p "Warm narration with a gentle smile." -v Achernar -o warm
gemgen tts -t "The glacier moved a few inches each day." -p "Calm documentary voice." -v Charon -e MP3 -o doc
gemgen tts -t "[whispering] The door is open." -p "Whispered warning." -v Kore -o warning
gemgen tts -t "[extremely fast] Terms apply. See store for details." -p "Fast disclaimer." -r 1.8 -o disclaimer
gemgen tts --speaker Sam=Kore --speaker Bob=Charon --turn "Sam:Did you hear that?" --turn "Bob:[laughing] I did." -p "Amused conversation between two friends." -o chat
gemgen tts --speaker Host=Achernar --speaker Guest=Puck --turn "Host:Welcome back." --turn "Guest:Good to be here." -p "Two-speaker dialogue, relaxed interview." -o interview
gemgen tts --json-template > request.json
gemgen tts -i request.json -o batch
gemgen tts -t "Support is available now." --profile telephony-class-application -e MULAW -o phone

temperature is Vertex-only in the Gemini-TTS docs and is not exposed in this public page/form v1 flow.

Values

Voices: Achernar, Achird, Algenib, Algieba, Alnilam, Aoede, Autonoe, Callirrhoe, Charon, Despina, Enceladus, Erinome, Fenrir, Gacrux, Iapetus, Kore, Laomedeia, Leda, Orus, Pulcherrima, Puck, Rasalgethi, Sadachbia, Sadaltager, Schedar, Sulafat, Umbriel, Vindemiatrix, Zephyr, Zubenelgenubi.

Languages: ar-EG, bn-BD, nl-NL, en-IN, en-US, fr-FR, de-DE, hi-IN, id-ID, it-IT, ja-JP, ko-KR, mr-IN, pl-PL, pt-BR, ro-RO, ru-RU, es-ES, ta-IN, te-IN, th-TH, tr-TR, uk-UA, vi-VN, af-ZA, sq-AL, am-ET, ar-001, hy-AM, az-AZ, eu-ES, be-BY, bg-BG, my-MM, ca-ES, ceb-PH, cmn-CN, cmn-tw, hr-HR, cs-CZ, da-DK, en-AU, en-GB, et-EE, fil-PH, fi-FI, fr-CA, gl-ES, ka-GE, el-GR, gu-IN, ht-HT, he-IL, hu-HU, is-IS, jv-JV, kn-IN, kok-IN, lo-LA, la-VA, lv-LV, lt-LT, lb-LU, mk-MK, mai-IN, mg-MG, ms-MY, ml-IN, mn-MN, ne-NP, nb-NO, nn-NO, or-IN, ps-AF, fa-IR, pt-PT, pa-IN, sr-RS, sd-IN, si-LK, sk-SK, sl-SI, es-419, es-MX, sw-KE, sv-SE, ur-PK.

Encodings: LINEAR16, ALAW, MULAW, MP3, OGG_OPUS, PCM. Output extensions: LINEAR16 -> .wav, ALAW -> .alaw, MULAW -> .mulaw, MP3 -> .mp3, OGG_OPUS -> .ogg, PCM -> .pcm.

Audio profiles: wearable-class-device, handset-class-device, headphone-class-device, small-bluetooth-speaker-class-device, medium-bluetooth-speaker-class-device, large-home-entertainment-class-device, large-automotive-class-device, telephony-class-application.

Markup tags: [sigh], [laughing], [uhm], [sarcasm], [robotic], [shouting], [whispering], [extremely fast], [scared], [curious], [bored], [short pause], [medium pause], [long pause].

Develop

deno task check
deno task lint
deno task test
deno publish --dry-run --allow-dirty
deno publish
deno run -A cli.ts --help
deno run -A cli.ts tts --help

Sources for option lists: Google Cloud Gemini-TTS docs and audio profile docs, checked May 1, 2026.