Article

Qwen3 TTS costs and voice limits on Together AI

For creators and app teams, check Together's Qwen3 TTS pricing, upfront credits, language support and caption limits before choosing a voice.

Editorial illustration for Qwen3 TTS costs and voice limits on Together AI: a document represents the research briefing. Not documentary evidence.

Together AI announced hosted Qwen3 TTS on November 20, 2026, giving creators and app teams a managed way to turn scripts into speech. It is an option for preset-voice narration or spoken app replies: the serverless rate is US$20 per million characters, or $2 for an illustrative 100,000 billable characters.

The main limits can decide the purchase before sound quality does: Hindi is not among the supported languages, and word timestamps are unavailable. A new self-service account also needs a $15 initial credit purchase; a $2 workload does not mean a $2 starting payment.

November 20 is the announcement date, not a settled first-availability date. Together’s model page also shows a September 28, 2026 release date, leaving the earlier availability history unclear.

What does speech generation cost?

At the current public serverless tariff, the calculation is: 100,000 billable characters ÷ 1,000,000 × $20 = $2. One million billable characters would cost $20. These are synthesis-only estimates, assuming one generation at list price, with no discounts, taxes or other service charges.

Use billable characters, not word count or minutes of finished audio. The cited pricing does not settle every character-accounting detail, including how retries or failed requests are charged; confirm those rules for your workload.

Together’s credit rules say there is currently no free trial. The first purchase must be at least $15 before playground use or API-key creation; later purchases start at $5. That $15 buys prepaid credit—it is not an extra fee attached to each batch. Access requires a positive balance, while Scale and Enterprise contracts follow their own terms.

For price context, the same pricing page lists Kokoro at $4 and Orpheus at $15 per million characters. Under the same 100,000-character assumption, that is $0.40 and $1.50 respectively. This compares tariffs, not voice quality, language coverage or equivalent functionality. The Qwen character rate is not a quote for dedicated capacity.

Which voices and languages can you use?

Together offers nine preset voices across ten languages: Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish. Each preset can speak all ten. Hindi is outside this documented list; do not choose this endpoint for a Hindi requirement without further confirmation.

Examples include Ryan and Aiden for native English voices, Ono_Anna for Japanese and Sohee for Korean. Qwen recommends native-language pairings for best quality, so multilingual support should not be read as equal performance in every pairing.

Supply the script in the intended language: Together’s multilingual examples use separately translated text. Changing the language setting is not a translation workflow.

The Qwen-specific guide permits delivery instructions of up to 500 characters to control tone, emotion or pace. That limit applies to the instruction field, not an established maximum length for the script being spoken.

Preset control is also different from cloning a particular person’s voice. Qwen’s model family separates CustomVoice, VoiceDesign and Base variants for preset delivery, description-based voice design and reference-audio cloning. Do not assume this hosted preset endpoint includes every family capability.

Will it deliver the audio and captions you need?

For saved narration, the Qwen-specific format list supports WAV, MP3, raw PCM and mu-law, at 24,000 or 8,000 Hz. Opus, AAC and FLAC are not supported. If an editor or playback system requires one of those formats, allow for conversion or choose another service.

Word timestamps—the timing needed to match individual words to the audio—are not supported for Qwen on Together. For synchronized captions, budget a separate alignment step or choose an endpoint that supplies the timing. Generating a voice track alone does not deliver timed subtitles.

For live playback, Together’s HTTP-streaming guide describes audio arriving in chunks before the full response is finished. It specifies raw PCM, with chunks encoded for transport and decoded by the application. Do not assume MP3 file support means incremental MP3 streaming. The general format table and dedicated streaming guide differ on mu-law streaming, so confirm that transport separately if it is essential.

What else does a voice app need?

Speech synthesis is the speaking component, not a complete assistant. Together’s phone-agent walkthrough combines transcription, a chat model, speech generation, telephony and application logic. Budget those components and hosting separately; the $2 example cannot establish a price per call.

That separation also matters in PolyAI Dialog-RSN-1 voice pilots on AWS, which need a speech-output engine. This is an architectural parallel, not evidence of a tested Qwen integration. For a different billing approach, see GPT-Live active and standby session costs; session-minute charges are not directly comparable with a synthesis-only character rate.

Together advertises first audio in as little as 97 milliseconds. That is a vendor claim, not an independently measured response time for your application. Listening, generating a reply, network delivery and playback add other stages.

The serverless rate-limit documentation describes performance as best-effort and warns of demand-related 429 or 503 errors. The product page’s general SLA wording does not establish a guarantee for every self-service account. Confirm applicable capacity and contractual coverage before promising a response time to customers.

What should you confirm before committing?

For a small evaluation, use the documented model identifier Qwen/Qwen3-TTS-12Hz-1.7B from the product page. Have speakers of your target languages assess representative names, numbers and sentences. Evaluate playback in the actual application, not just a downloaded sample.

Check commercial terms separately from the model licence. The CustomVoice model card identifies Apache-2.0, while Together’s service terms govern hosted use. Those terms allocate input and output ownership to the customer as between the parties, but retain content responsibilities and model-specific conditions; they do not clear someone else’s script for reuse.

Before a larger rollout, confirm your account’s concurrency allowance and maximum speech-text length; neither is established by the cited Qwen guide. The practical fit is preset narration or app speech where the supported languages, delivery format and surrounding components meet the task. If Hindi, built-in caption timing or cloning is a must-have, settle that requirement before committing to this endpoint.