TTS voices: Qwen and Gemini specs, voices, pricing

Playcut’s voice surface has three parts: text-to-speech (turn text into an audio asset), voice design (describe a voice in words and get a reusable voice), and voice cloning (upload a short sample and get a voice that sounds like it). TTS is billed per 1,000 characters and the rate depends on which provider the voice you picked belongs to — Qwen 15 credits, Gemini 25 credits. Qwen is the default and the cheapest, and it is the only provider that can design or clone voices. Gemini is the premium tier: the widest language coverage and the strongest natural-language style steering. Voice design costs a flat 20 credits; cloning from your own audio is priced at 5 credits.

Specs at a glance

	Qwen	Gemini
Provider	Alibaba (DashScope)	Google
Status	Live — default	Live
Credits / 1,000 chars	15	25
Text-to-speech	✅	✅
Voice design	✅	❌
Voice cloning	✅	❌
Max input (Playcut API)	600 chars (pre-hold enforced)	5,000 chars
Max input (provider)	600 chars	32k-token session
Style steering	free-text instructions (EN/ZH only)	inline [tags] + leading directive
Output audio	container follows response content-type; stamped 24 kHz mono	WAV, 24 kHz, mono, 16-bit

Aspect ratio, resolution, and duration are all N/A for voice — output length is a function of the input text. There is no duration parameter on any TTS path.

Qwen requests are capped at 600 characters

Qwen’s provider-side input cap is 600 characters per request. Playcut enforces this before any credits are held: a longer request to a Qwen voice returns a clear 400 error, and the character counter in the app shows the 600 limit when a Qwen voice is selected. Gemini voices accept up to the 5,000-character API cap. For long scripts on Qwen, split the text and generate each part separately.

Model ids behind each provider

Qwen carries three distinct model ids — one for synthesis, one for design, one for clones. You never pass these directly; Playcut resolves them from the voice you select.

Provider	Model id	Used for
qwen	qwen3-tts-instruct-flash-2026-01-26	Text-to-speech with built-in voices
qwen	qwen3-tts-vd-2026-01-26	Voice design
qwen	qwen3-tts-vc-2026-01-22	Voice cloning, and playback of designed/cloned voices
gemini	gemini-3.1-flash-tts-preview	Text-to-speech with prebuilt voices

Modes

Text-to-speech

• Tool: generate-tts
• Requires: text + voiceId — up to 600 chars for Qwen voices, 5,000 for Gemini (checked before credits are held)
• Optional: workerId (session), language, prompt (style direction, ≤200 chars)
• Cost: 15 / 20 / 25 credits per 1,000 characters, depending on the voice’s provider. Billed as max(1, ceil(length / 1000)) units — a 12-character request still costs one full unit.

The provider is determined by the voice, not by a parameter. Fetch your voice library with list-voices and read each voice’s provider before you budget.

Voice design

Describe a voice in natural language and get a reusable voice back — no audio sample required. Qwen only; Gemini has no design API.

• Tool: design-voice
• Requires: prompt (the voice description)
• Optional: text (preview script), language, voiceName
• Cost: 20 credits flat

Voice cloning from audio

Upload a short recording and Playcut enrolls it as a voice you can synthesize with. Qwen only.

• Tool: clone-voice-from-audio (takes an audioAssetId) · REST alternative: POST /api/voices/clone-from-audio (multipart)
• Requires: name (1–100 chars) + the audio
• Optional: language, referenceText (≤500 chars)
• Listed price: 5 credits — but see the note in Pricing; this is currently charged at 0.

Provider requirements for the sample: at least 5 seconds of continuous clear speech (10–20 seconds recommended), 60 seconds maximum, ≤10 MB, in WAV (16-bit), MP3, or M4A. Playcut validates size and duration before any credits are held (a bad sample is a free, instant 400), and automatically converts other containers (WebM mic recordings, OGG, FLAC) to 24 kHz WAV so they just work.

Voices

Gemini — 30 prebuilt voices

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

These are cross-lingual — the same voice name speaks any supported language. They are seeded into every workspace’s built-in library.

Qwen — built-in voices

Alibaba documents 41 voices across realtime and non-realtime models. The English-relevant set Playcut seeds is Cherry, Ethan, Serena, Moon, Vivian, Kai, Maia, Vincent, Neil, Mochi. Not every documented Qwen voice works on the instruct-flash model Playcut uses for synthesis — a handful (Ryan, Eric, Dylan, Aiden, Sohee) exist on a sibling model and are deliberately excluded because they fail on ours.

Languages

Provider	Languages	How language is selected
Qwen	11 values: Auto, Chinese, English, German, Italian, Portuguese, Spanish, Japanese, Korean, French, Russian	language parameter (language_type upstream)
Gemini	90+ BCP-47 entries published by Google	Auto-detected from the text. There is no language parameter — Playcut accepts language for API symmetry and ignores it

Qwen’s regional Chinese dialects (Shanghainese, Cantonese, Sichuan, Beijing, Nanjing, Shaanxi, Southern Min, Tianjin) are carried by specific dialect voices upstream, not by the language parameter. Playcut exposes none of them.

Style control

Gemini takes plain-text direction two ways, and there is no closed vocabulary — the tags are open-ended natural language:

[excitedly] Hey there, I'm a new text to speech model.
[sarcastically, one painfully slow word at a time] Hey there.

A leading directive works too (Say cheerfully: Have a wonderful day!); Playcut builds it from the optional prompt field, which is capped at 200 characters.

Qwen takes free-text instructions on the instruct-flash model only, and Alibaba documents that field as Chinese and English only. A style prompt written in Spanish is outside the supported envelope even when the spoken text is Spanish.

Cloned and designed voices ignore style direction entirely. Both play back through Qwen’s voice-clone model, which drops the instruction field by design to preserve the speaker’s acoustic identity. If you need a whisper from a cloned voice, clone from audio that already contains one.

Pricing

Operation	Playcut credits	Provider cost	Notes
TTS — Qwen	15 / 1,000 chars	~$13 / 1M chars — unverified	Default provider. Billed max(1, ceil(len/1000))
TTS — Gemini	25 / 1,000 chars	$1.00 / 1M input tokens + $20.00 / 1M audio output tokens	Our priciest TTS
Voice design	20 flat	~$0.20 per design (unverified) + $0.01 mandatory re-enrollment ≈ $0.21	Includes the save/clone step
Voice clone from audio	5	$0.01 per voice	Charged at submission, refunded on failure

Three honest pricing caveats

• Qwen’s provider cost is unverified. qwen3-tts does not appear on Alibaba’s Model Studio pricing page; the $13/1M figure comes from an internal snapshot, not a live source. It does not affect what you pay — 15 credits per 1,000 characters is the number that bills.
• Gemini’s per-character cost is our own derivation. Google publishes token rates, not a per-character or per-minute price. The ~$40/1M-chars figure used internally is arithmetic, not a Google number.
• Voice cloning from audio is currently free. The pricing table says 5 credits, and the pre-flight affordability check uses 5, but the actual credit hold resolves to 0. Treat 5 as the number that may start applying, and 0 as what you are charged today.

Limitations & gotchas

• Voice-clone samples are validated up front (since 2026-07-23): samples over 10 MB, shorter than 5 seconds, or longer than 60 seconds are rejected with a clear message before any credits are held. OGG/FLAC/WebM uploads are converted to WAV automatically instead of failing at the provider. Sample-rate is not checked — the provider remains the backstop there.
• A “designed” voice is physically a cloned voice. Qwen’s design call returns a voice id Playcut cannot synthesize with over its REST path, so the preview clip is re-enrolled through the clone model and that id is what is stored. Two consequences: designed voices cannot be style-steered, and the preview clip is written to a public storage bucket so the provider can fetch it.
• Cross-lingual support is asserted, not measured. Every designed and cloned Qwen voice is stamped with all 10 cross-lingual languages regardless of the language it was created in. A voice designed in English is claimed to speak Korean; nothing verifies the quality.
• Gemini output format is fixed. Always WAV, 24 kHz, mono, 16-bit. Google exposes no encoding, sample-rate, speed, pitch, or volume parameters on Gemini TTS (those belong to classic Cloud TTS neural voices). Steer delivery with [slowly] / [shouting] style tags instead. The app’s former “Advanced” encoding panel was removed 2026-07-21 — it never controlled anything.
• temperature, topK, and topP were removed (July 2026). No synthesis backend ever consumed them; sending them now returns a 400 instead of being silently dropped.
• Gemini style direction is prepended to your text. A badly worded instruction can be read aloud instead of applied. The 200-character cap on prompt limits the blast radius but does not prevent it.
• Gemini is a preview model with no dated snapshot. It is published only under the unsuffixed -preview id, so behavior can change without an id change, and preview rate limits are tighter.
• Qwen’s output sample rate is assumed, not measured. Playcut stamps every Qwen asset as 24 kHz mono without inspecting the returned audio. Alibaba does not formally document the container or rate for the non-streaming HTTP path.
• Custom voice quota is untracked. Alibaba allows up to 1,000 custom voices per model family per account, and auto-removes voices unused for a year. Playcut neither enforces nor garbage-collects against either limit.
• Rate limits are shared. Qwen TTS shares a per-process DashScope request window with Wan 2.7 and HappyHorse video; Gemini TTS shares one with Gemini image generation, Omni video, and actor interactions. Heavy video load throttles voice.

How to generate

Text-to-speech

curl -X POST https://api.playcut.ai/api/v1/tools/generate-tts/execute \
  -H "Authorization: Bearer $PLAYCUT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Welcome to Playcut. Generated entirely from a single API call.",
    "voiceId": "<voice-id-from-list-voices>",
    "prompt": "warm, unhurried, like a late-night radio host"
  }'

Find a voiceId first:

curl -X POST https://api.playcut.ai/api/v1/tools/list-voices/execute \
  -H "Authorization: Bearer $PLAYCUT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{}'

Design a voice

curl -X POST https://api.playcut.ai/api/v1/tools/design-voice/execute \
  -H "Authorization: Bearer $PLAYCUT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "Mid-30s warm female, soft Boston accent, comforting podcast-host energy",
    "voiceName": "Harbor",
    "language": "English",
    "text": "This is what my voice sounds like."
  }'

Clone from audio

curl -X POST https://api.playcut.ai/api/v1/tools/clone-voice-from-audio/execute \
  -H "Authorization: Bearer $PLAYCUT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "audioAssetId": "65f3a1b2c3d4e5f600000003",
    "name": "Narrator (me)",
    "language": "English"
  }'

Upload the sample first so you have an audioAssetId. Keep it between 10 and 20 seconds, WAV/MP3/M4A, at least 24 kHz.

MCP

With the Playcut MCP server installed in Claude Desktop, Claude Code, or Cursor, call the same tools by name:

• generate-tts
• design-voice
• clone-voice-from-audio
• list-voices

See the MCP install guide.

All three return an async acknowledgement. Poll with get-task-status, then fetch a signed URL with get-asset. Full flow in the Quickstart.

{
  "data": {
    "taskId": "65f3a1b2c3d4e5f600000002",
    "status": "PENDING",
    "estimatedCredits": 15
  }
}

Tip

estimatedCredits on generate-tts is computed at the default provider rate (15 per 1,000 chars). If your voice belongs to Gemini, the actual charge resolves from the voice document and will be higher. Check the provider on the voice before you rely on the estimate.

Related

• All models — the full catalog index, side by side
• Resolutions and pricing — every model’s credit price side by side
• Lyria 3 — the other audio surface: text-to-music, flat 20/40 credits per song
• Veo 3.1 — video that generates its own synchronized dialogue, no TTS pass needed
• Aspect ratios by model — the cross-model ratio table (image and video)
• Quickstart — end-to-end generation in five steps
• API Reference — the full tool list and live Swagger UI