Skip to content
trendoviz

Together AI adds Qwen3 TTS with 97 ms first audio

Importance: NotableStatus: Developing Provisional Confidence: Primary-source claim

Together AI has made Qwen3 TTS available through its Serverless Inference and dedicated endpoints, offering nine preset voices, plain-language delivery instructions and support for 10 languages. The vendor says first audio can arrive in as little as 97 ms at a listed price of $20 per million characters.

52 out of 100Emerging7D: Insufficient history ProvisionalPublished

Why it is important now

Importance: Notable Confidence: Primary-source claimStatus: Developing ProvisionalModel or API release
First detected:
Event / publication time:
Last updated:
Observed for:
0 hours

Provisional

  • Not yet corroborated or confirmed by an independent source.
  • First detected less than 48 hours ago.
  • At least one central claim is not yet supported by independent or authoritative evidence.

Why it is important

  • Rated consequence medium, immediacy medium.
  • The responsible organisation published this itself.
Importance dimensions
DimensionRatingBasis
ScopeUnknown (not counted against it)—
ConsequenceMediumconsequence.type_model_release
ImmediacyMediumimmediacy.type_model_release
AuthorityHighauthority.own_products
ActionabilityMediumactionability.review_or_migration
IrreversibilityUnknown (not counted against it)—

Rule dimensions.notable · events-1.0+importance-1.0+general-1.0. Importance is not popularity: coverage volume is not an input.

Evidence confidence

Stated by the responsible organisation itself. Authentic, but not independently verified.

  • The responsible organisation published this itself.

What is known

Primary-source or vendor statement

  • Together AI blog: “Qwen3 TTS is now available on Together AI”[1] (not yet verified · primary source claim)

Sources behind this assessment

  1. [1] Qwen3 TTS is now available on Together AI — Together AI blog (primary source) · published Nov 20 · retrieved Nov 20

Classification history

  1. · status: Developing (rule, events-1.0) — Rated consequence medium, immediacy medium.

Overview

Together AI's November 20, 2026 announcement describes Qwen3 TTS as an open text-to-speech model available on both Serverless Inference and dedicated endpoints [1]. The documented model identifier is Qwen/Qwen3-TTS-12Hz-1.7B, and requests can be made through Together's Python and TypeScript SDKs or with cURL, using parameters for voice, language and delivery instruction [1].

The model is organized around nine preset voices. A voice stays the same while an instruction such as "excited and upbeat" changes how the text is delivered, and each preset supports Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian [1]. Together AI says it tuned its serving stack for the model, with first audio reported in as little as 97 ms [1].

Two streaming paths are documented. For complete text, the HTTP streaming mode returns server-sent events whose conversation.item.audio_output.delta events carry base64-encoded audio chunks [1]. Together AI also references a WebSocket API for text that arrives in pieces, though the announcement excerpt does not explain its implementation [1].

Why it matters

Keeping one recognizable voice while varying its delivery can simplify product design for multilingual assistants and other spoken interfaces, since teams do not need a separate preset for each speaking style [1]. HTTP streaming suits applications that already have the full text, while the referenced WebSocket option is intended for text generated incrementally, a pattern relevant to voice agents [1].

VoiceFinder gives teams a way to compare presets on their own copy and export matching request code, which may shorten the initial selection and prototyping work [1].

Opportunities

Potential uses include multilingual assistants, voice agents and interactive products that want a consistent voice across several markets [1]. Teams can use VoiceFinder to test voices against their own lines before committing to a preset, then reuse the generated code as a starting point for an integration [1].

The HTTP streaming format is documented in enough detail to help teams plan applications in which speech generation begins before an entire response is ready [1].

Risks

A misspelled instructions field is skipped without raising an error. If an application names the parameter incorrectly, audio may be returned with default delivery rather than the intended tone, making field validation and output checks necessary [1].

Cross-language support does not mean uniform quality. Qwen recommends using each voice in its native language, so applications covering several languages may need per-language testing [1]. The absence of word-level timestamps may also affect products that depend on precise word synchronization or timing [1].

The announced price of $20 per million characters applies to Serverless Inference and dedicated endpoints as presented, but the source does not distinguish pricing between those endpoint types, so total project cost cannot be inferred from that figure alone [1].

Context

Qwen3 TTS separates speaker identity from delivery: developers select a preset voice, specify a language and optionally describe the intended tone. The announcement demonstrates the same English voice reading the same line with different instructions, and reading a translated order update in four languages [1].

Each voice also has a native language, and Qwen recommends using a voice in its native language for best quality. Together AI therefore advises listening to the languages an application needs before settling on a preset [1].

For voice selection, Together AI's VoiceFinder lists the Qwen3 TTS voices, plays a sample of each, reads custom text of up to 500 characters and shows the corresponding request code in Python, TypeScript or cURL [1]. Word-level timestamps are not supported: alignment=word is explicitly excluded for this model [1].

Known, claimed, inferred

Availability through both endpoint types, the model identifier, the request methods, the HTTP streaming event format, the lack of word-timestamp support and VoiceFinder's described behavior are documented facts [1]. The 97 ms first-audio figure, the price, the serving-stack tuning and the model's language and voice-consistency characteristics are vendor statements rather than independently validated results [1].

The availability of HTTP streaming alongside a WebSocket interface indicates that developers can choose an access pattern based on when text becomes available. However, the announcement excerpt references the WebSocket API without documenting its request or response behavior, so teams would need fuller technical documentation before evaluating that path [1].

Limitations & uncertainties

There is no independent or third-party evaluation of the claimed 97 ms first-audio latency [1]. The source provides no audio-quality measurements, human ratings or comparisons with other text-to-speech models [1].

The announcement excerpt does not detail the WebSocket API, and it does not list the nine preset voice names or their individual native languages, although it notes that each voice has one [1]. Dedicated-endpoint pricing is not distinguished from Serverless Inference pricing, and no licensing terms for the open Qwen3 TTS model are provided [1]. The account rests on a single vendor source, so there are no cross-source comparisons or disagreements to assess [1].

Evidence & sources

The key statements of this analysis, labeled by type, with the sources they rely on. Trendoviz does not run product tests.

  • Primary-source or vendor statementTogether AI made Qwen3 TTS available on Serverless Inference and dedicated endpoints on November 20, 2026.[1]
  • Primary-source or vendor statementThe documented model identifier is Qwen/Qwen3-TTS-12Hz-1.7B, with request examples for Python, TypeScript and cURL.[1]
  • Primary-source or vendor statementTogether AI lists Qwen3 TTS at $20 per 1 million characters.[1]
  • Primary-source or vendor statementTogether AI says first audio can arrive in as little as 97 ms after its inference team tuned the serving stack for Qwen3 TTS.[1]
  • Primary-source or vendor statementThe model offers nine preset voices whose delivery can be changed with plain-language instructions while the voice itself remains the same.[1]
  • Primary-source or vendor statementEvery preset voice is described as supporting Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian.[1]
  • Primary-source or vendor statementFor full-text requests, the HTTP streaming mode returns server-sent events with base64-encoded audio chunks in audio_output.delta events; a WebSocket API is referenced for text arriving in pieces.[1]
  • Primary-source or vendor statementWord-level timestamps using alignment=word are not supported for Qwen3 TTS.[1]

Sources

  1. 1
    Qwen3 TTS is now available on Together AI
    ●together.airetrieved Nov 20, 2026

Updates & corrections

  1. First published · published automatically after passing every check

Found an error? Report it.