Qwen-Audio 3.1 expands Alibaba’s voice stack and cuts API prices

Alibaba’s Qwen team has released Qwen-Audio 3.1, an update spanning automatic speech recognition, text-to-speech, and full-duplex conversation. The release upgrades the existing ASR, TTS, and Realtime models and introduces ASR-Next for richer audio analysis and TTS-Next for long-form productions that combine speech, sound effects, and background audio.

Alibaba reports substantial reductions from its previous API prices:

• Text-to-speech: approximately 70%.
• Realtime conversation: approximately 85%.
• Speech recognition: up to 95%.

Voice applications can source transcription, generation, and conversational turn-taking from one hosted lineup. That can reduce integration work, although latency, quality, and vendor risk still require testing at the endpoint level.

Five models cover the voice loop

TTS runs on fewer speech tokens

The accompanying technical paper provides the most detail about Qwen-Audio 3.1 TTS. Its low-frame-rate tokenizer represents speech at 12.5 frames per second, reducing the number of units that the autoregressive language model must predict. A flow-matching component then generates the acoustic detail needed for the final waveform.

The five-stage training process separates early specialization from later joint optimization:

1. Pretrain the language model.
2. Pretrain the flow-matching model.
3. Train both components jointly while concentrating progressively on higher-quality data.
4. Apply reinforcement learning to the language model.
5. Apply reinforcement learning to the flow-matching model.

Alibaba says the reinforcement-learning stages improve prosody, voice similarity, and resilience to difficult reference recordings. The model accepts free-form instructions for role, emotion, speaking rate, timbre, style, and accent. It also supports 86 inline tags for nonverbal events such as laughter, breathing, coughing, and sighing.

Language coverage extends to 16 languages and 20 Chinese dialect regions. The model can synthesize as much as three minutes of audio in one pass and accepts reference speech containing noise, reverberation, or unclear pronunciation.

Realtime collapses three handoffs

Qwen-Audio 3.1 Realtime Plus expands the context window to 262,144 tokens and supports full-duplex audio, function calling, web search, and voice cloning. It accepts audio and text while streaming speech and text responses over persistent realtime connections.

Conventional voice agents commonly route each turn through speech → ASR → text → language model → text → TTS → speech. The integrated model folds speech understanding, reasoning, and generation into one service. Fewer orchestration boundaries can reduce latency and simplify interruption handling, especially when a user speaks over the agent.

Applications already using Qwen-Audio 3.0 Realtime Plus can retain the same integration protocol, which should limit code changes during migration. Production teams will still need regression tests for turn-taking, tool calls, output quality, and error handling.

The 262,144-token limit serves as a ceiling for long sessions rather than a target for retained history. Live audio, tool results, and conversation state all consume context, so stateful agents still need summarization, truncation, and latency controls as sessions grow.

A top ranking, with limits

In July 2026, the previous Qwen-Audio 3.0 TTS Plus model ranked first in the Artificial Analysis Text-to-Speech arena with a Quality Elo score of 1,237. Elo summarizes comparative preference results, with higher scores indicating stronger evaluator preference. Qwen ranked ahead of Google Gemini 3.1 Flash TTS, MiniMax Speech 2.8 HD, and ElevenLabs Eleven v3.

The prior model listed at $27.60 per million characters, about one-third of the listed rates for the compared ElevenLabs and MiniMax tiers. That result supplies historical context for the 3.1 line; the new endpoints need separate quality and latency measurements.

The published comparison also records two constraints:

• Throughput: Qwen-Audio 3.0 TTS Plus generated about 16 characters per second, compared with 30.2 for Simba 3.2, 27 for Gemini 3.1 Flash TTS, and 120 for Sonic 3.5.
• Statistical confidence: Qwen’s lead over Simba 3.2 fell within overlapping confidence intervals, making the top two results statistically tied.

Hosted access comes first

Alibaba currently exposes the announced 3.1 endpoints through Alibaba Cloud Model Studio:

The open-weight Qwen3-TTS repository belongs to a separate lineage from the hosted Qwen-Audio 3.x services, so developers should verify feature and API parity before planning a migration. Alibaba has not provided launch dates for the ASR-Next and TTS-Next APIs.

Three practical fits

1. Voice agents and phone automation. Realtime Plus targets support calls, booking systems, and other long-running conversations that need interruption handling, tool use, and persistent state. The larger context window raises the session ceiling, while production memory policies remain necessary.
2. Audiobooks, dubbing, and podcasts. TTS Plus targets narration and controlled delivery. TTS-Next extends that workflow to productions requiring generated speech, effects, and background sound in the same pass.
3. Meetings and media analysis. ASR-Next combines speaker labels, timestamps, emotion tags, and non-speech event recognition in one model call, reducing the number of specialized services needed for searchable recordings.

Benchmarks to run before switching

Production trials should measure the behavior that aggregate rankings and vendor specifications leave unresolved:

• Turn latency: Record time to first audio, interruption cutoff time, and p95 response latency in each deployment region.
• Target-language quality: Measure transcription errors, pronunciation, prosody, and voice similarity using representative accents, dialects, and noisy recordings.
• Transcript fidelity: Confirm how the ASR polishing pass handles fillers and repetitions when applications require verbatim records.
• Tool reliability: Test function-call accuracy during interruptions, overlapping speech, and long conversations.
• Context growth: Track token consumption from audio, text, and tool results, then define summarization and truncation policies.
• Operating constraints: Verify concurrency limits, rate limits, data residency, retention rules, and voice-cloning safeguards.
• Total cost: Model the provider’s billing units against actual call duration, generated characters, retries, and peak traffic.

Teams can replay current production traces against the 3.1 endpoints and compare them with existing providers on latency, quality, reliability, and workload cost. Those measurements will determine whether the integrated model can replace a modular speech pipeline.