Alibaba just dropped speech-to-text through the price floor — and changed the meter on the way down

What shipped

On 22 September 2026, at Alibaba Cloud’s Apsara Conference, the Qwen team replaced its audio line with five models under one version number.

Three are upgrades. ASR widens dialect coverage to 30 languages and 16 Chinese dialects and adds native transcript polishing — filler and repetition removal at the model layer rather than in your post-processing — with first-character response around 160 milliseconds. TTS adds cross-lingual timbre transfer, holding one voice’s identity across Mandarin, Cantonese, English and Japanese, and replaces version 3.0’s 86 inline delivery tags with natural-language instructions for emotion, speed and style. Realtime goes full-duplex: speaking and listening at once, interruptible at any point.

Two are new. ASR-Next does multi-speaker recognition with speaker labels, timestamps and aligned transcripts, and extends past speech into emotion, ambient sound and machine-noise understanding — sound captioning, event localisation, audio question-answering. TTS-Next runs a combined language-model-and-diffusion framework that generates dialogue, sound effects and background audio in a single pass, aimed at audiobooks, podcasts, games and ads.

Alongside them Alibaba cut simultaneous-interpretation latency in Qwen3.8-LiveTranslate by nearly 20%, from 2.8 to 2.3 seconds.

That is a complete audio stack, announced competently. It is also not the reason this release matters.

The arithmetic nobody ran

The coverage has settled on three percentages: TTS down about 70%, Realtime about 85%, ASR up to 95%. Percentages are not prices. Here is the conversion.

Alibaba’s Beijing-node rate for Qwen-Audio-3.1-ASR-Flash is ¥0.8 per million input tokens and ¥2.7 per million output tokens. At the 23 September 2026 rate of 6.71 yuan to the dollar, that is $0.119 and $0.403.

Qwen’s documented ASR tokenisation is 25 audio tokens per second. An audio-hour is therefore 3,600 × 25 = 90,000 input tokens, costing $0.0107. An hour of speech transcribes to roughly 9,000 words — call it 12,000 output tokens — adding $0.0048.

Roughly $0.016 per audio-hour.

For scale, our analysis of Gemini 3.5 Transcribe in August mapped the cheap tier of this market, and it had barely moved in years:

The thesis of that August piece was that Google had walked down to a floor the specialists set years ago, not reset it. That reading was right then and it is what makes this different now. Qwen-Audio-3.1-ASR is roughly eleven times below Meta’s $0.18 and fourteen times below Gemini’s $0.30. This is the first move in this category that is a break rather than a convergence.

The meter moved on the same day

Now the part that should slow you down.

The predecessor, qwen3-asr-flash, billed by audio duration — $0.000035 per second on the international deployment, or $0.126 per audio-hour. Version 3.1 bills per token. On the synthesis side, Qwen-Audio-3.0-TTS billed per character; 3.1 bills per token.

Against that $0.126 duration price, our $0.016 works out to an 88% cut — real, large, and not the 95% on the banner. The gap is not necessarily spin; the 95% may describe a different tier or the domestic rate card. But it cannot be checked, because the unit changed underneath the comparison and Alibaba did not republish the audio-token conversion rate alongside the new prices. The 25-tokens-per-second figure is documented for Qwen ASR generally, not restated for 3.1.

This is the second time this month a voice vendor has repriced by changing what it counts. When GPT-Live-1 went to a flat per-minute meter on 10 September, the headline was a price cut and the substance was that silence stopped being free — hold queues and thinking pauses became billable. Qwen has run the same play in the opposite direction, and the consequence is the mirror image.

Duration and character billing are quotes. An hour is an hour; a million characters is a million characters. Token billing is an estimate, split in two. The audio input side stays fixed, because audio tokenises at a constant rate no matter what is on the recording. The output side moves with content: a dense four-person technical panel generates far more transcript per audio-hour than a one-on-one with long pauses. On the numbers above, that variable component is about 30% of the blended cost.

That is precisely the structural criticism we levelled at Gemini 3.5 Transcribe’s token billing in August — now adopted by the vendor undercutting it. At a tenth of the price the variance matters far less in absolute terms. It still means your forecast is a model, not a quote.

There is one encouraging cross-check. Run the same arithmetic on Qwen-Audio-3.1-TTS-Flash at ¥1.5 input and ¥12 output per million tokens and you get roughly $3.30 per million synthesised characters, against about $15 for the 3.0 Flash tier it replaces — a 78% cut, close enough to Alibaba’s claimed ~70% that the tokenisation assumption appears sound. Two independent derivations landing near the vendor’s own figures is the best verification available without a published conversion table.

What you cannot buy yet

Four things stand between this rate card and a migration.

Two of the five models have no API. ASR-Next and TTS-Next — the multi-speaker recognition and the single-pass audio-scene generation, which are the genuinely novel entries — were still listed as pending at announcement. They are demos.

The rate card is not in your region. Every yuan figure above is the Beijing node. The international pay-as-you-go path runs through the Singapore-anchored ap-southeast-1 endpoint in USD at different rates — realtime transcription there moved to $0.93 per million input and $0.70 per million output tokens. At the time of writing, Alibaba Cloud’s international Model Studio pricing documentation did not list the Qwen-Audio-3.1 family at all. You cannot sign a contract against a page that does not exist.

Beijing is a data-residency answer, not a location detail. Call recordings and meeting audio are among the most sensitive data most organisations hold. For an EU or US buyer this is a GDPR and procurement question that precedes the cost question entirely — the same split we saw when Grok 4.6 landed on Bedrock with a data-residency premium, where the cheap endpoint and the compliant endpoint were not the same endpoint.

Quality is unmeasured. Qwen-Audio-3.1 has no independent score. The verified 1,259 Elo on the Artificial Analysis Provider Voice Arena belongs to the 3.0 Plus tier — the model this release supersedes. A price an order of magnitude below the market is only interesting if the output is usable, and right now the only evidence for that is the vendor’s own.

The synthesis side, and why price is not the argument

For text-to-speech the gap is wider still and the case is weaker.

At roughly $3.30 per million characters, Qwen sits against ElevenLabs Eleven v3 at about $100 per million characters as tracked by Artificial Analysis, with Flash near $50. A 20-million-character monthly narration workload is roughly $2,000 on Eleven v3 and roughly $66 on Qwen.

A 30x gap will get a procurement meeting. It should not win it on its own, because characters are not what you are buying.

The day after Qwen’s announcement, Google shipped Gemini 3.8 Flash TTS with consent-gated 30-second voice replication, SynthID and C2PA provenance marking — and blocked voice replication in the EU outright rather than test Article 50. Qwen’s announcement says nothing about provenance marking at all. For anyone publishing synthetic audio into Europe, that silence is a compliance exposure that no per-character saving offsets, and it is the same gap that made Lyria 3.5’s SynthID marking the deciding factor in music generation rather than output quality.

There is also a licence-posture question. Qwen has changed terms under shipped products before — Qwen-Image 2.1 moved from Apache 2.0 to research-only on 21 September, a month after launch. That is not a reason to avoid Qwen; it is a reason to keep your audio layer behind an abstraction you can repoint.

What to actually do

If you call a speech-to-text API directly, this is worth a bounded pilot — but on the international endpoint, with your own densest audio, measuring blended cost per audio-hour rather than trusting the per-token rate. The number to beat is whatever you pay now; if you are on a legacy enterprise stack like Google Cloud STT or AWS Transcribe, the case was already overwhelming before this release and is now absurd.

If you are on a specialist — Deepgram, AssemblyAI — price is now genuinely a reason to look, which it was not in August. It is still not a reason to move until residency, provenance and an international rate card exist. Use it in your renewal.

If you synthesise voice for publication in the EU, stay where provenance marking is documented. The gap is real and it is not the one on the invoice.

If you buy meeting or podcast software rather than build it — Otter.ai, Descript, and the rest of the audio tooling shortlist — do nothing. This is an API-layer release. Your vendor’s input costs just fell by an order of magnitude; you will experience that as slightly better transcripts and, eventually, as pricing pressure someone else has to apply.

The broader pattern is the one we have been tracking since the price war flipped in August: Chinese labs set the floor, Western labs walk down to it, and the floor is increasingly set at a level that looks less like a margin and more like a customer-acquisition budget. What is new here is that for the first time in speech, the floor did not hold. Developers building on audio should plan for a market where recognition is effectively free within two years — and where the thing you are actually paying for is the contract, the region and the provenance trail, not the tokens.

Prices converted at 6.71 CNY/USD as of 23 September 2026 and will move. Per-audio-hour figures are derived from Qwen’s documented 25-audio-tokens-per-second rate and an assumed 9,000 words per hour of speech; Alibaba has not published a conversion table for Qwen-Audio-3.1. Verify against your own audio before committing budget.

Update, 25 September 2026 — the other way to move a price is to delete it

This article’s thesis is that the meter change is the story, because a percentage quoted across a changed unit cannot be checked against your old invoice. Two days later Google demonstrated the limit case.

On 24 September, Gemini 3.8 Live with Live Avatar reached general availability with real-time generated video attached to Google’s speech-to-speech models — and no published rate for the video stream. The conversation bills on the existing Live API audio meter ($3 and $12 per million tokens, roughly $0.011 a conversational minute at Google’s own 25-audio-tokens-per-second rate — the same tokenisation constant used in the Qwen arithmetic above). The avatar itself is presented as a feature, not a product.

For the comparison shopper this is harder to handle than Alibaba’s 95%. A changed unit at least leaves you a number to convert; a missing unit leaves nothing to forecast against, and nothing preventing a rate appearing later. HeyGen charges $0.10 a minute in LITE mode for the equivalent layer, which is the market’s current opinion of what it is worth. That figure, not zero, is the number to put in a budget.

Questions or corrections? Email Pick Right. Want the full list? See all news.