Alibaba Cloud Model Studio ships several realtime speech models: Fun-ASR-Realtime for general streaming transcription, Paraformer-realtime-v2 as a legacy option, and Gummy-realtime-v1 when you want recognition and translation on the same stream. List prices in the China (Beijing) region run from about $0.08 to $0.17 per hour of audio for transcription alone. The LecSync Realtime Speech API is $0.006 per audio minute ($0.36 per hour) and returns captions, a paired translation across 60 languages, speaker labels and running minutes on one WebSocket.
So the choice is less "whose ASR is cheaper in China" and more "what has to ride on the live stream."
Everything below comes from public pages read on October 15, 2026. For Alibaba: the speech-to-text model overview, fun-asr-realtime (English USD list and Chinese CNY list), paraformer-realtime-v2, and the Gummy-realtime-v1 model card / realtime WebSocket docs (CNY list; STT and translation billed separately). For LecSync: the developers page, pricing docs and connect docs. USD figures for CNY-only rows use the Fun-ASR dual listing (CNY 0.00033/s and USD 0.000047/s), which implies about 7.02 CNY per USD.
The short answer
For Chinese-centric streaming transcription inside Alibaba Cloud's Beijing (or Singapore-routed) footprint, Fun-ASR and Paraformer are hard to beat on list price. Fun-ASR-Realtime is USD 0.000047 per second of audio — about $0.17 per hour — with hot words, dialect coverage and a 1,200 RPM default. Paraformer-realtime-v2 is cheaper still on the CNY card (CNY 0.00024/s ≈ $0.12/hr), though Alibaba's own overview now says to migrate toward Fun-ASR or Qwen-ASR when you can.
For live captions that someone reads in another language, with speakers and minutes on the same bill, LecSync is the flatter path. Translation, speaker labels, glossary and live digest are inside $0.36/hr. On Alibaba you either stay inside Gummy's language matrix and pay recognition plus translation as two meters, or you glue Fun-ASR to a separate MT service.
How each session works
LecSync. Your server calls POST https://api.lecsync.com/v1/realtime/connect with the audio format, a language pair and the features you want. It gets back a stream URL and a one-time token. The client pushes 16 kHz mono PCM. Transcript segments, their translations (matched by segment ID), speaker labels with millisecond timestamps (speakers: true) and sections of minutes (digest: true) all come back on that one connection. A network stall is capped at 2 seconds of backlog instead of letting latency climb.
Fun-ASR-Realtime / Paraformer. You open a DashScope WebSocket (Beijing or a Singapore workspace URL for International scope), send run-task with the model id, then stream binary audio. Results arrive as intermediate and final sentences with word-level timestamps. Hot words and VAD silence thresholds are first-class. The overview marks speaker diarization as unsupported on these realtime models; diarization is listed on offline / file-transcription models instead. Neither Fun-ASR-Realtime nor Paraformer-realtime-v2 returns a translation on the recognition stream.
Gummy-realtime-v1. One realtime model can turn on transcription, translation, or both. The WebSocket docs state that recognition and translation are billed separately at the same unit price, and that you must enable at least one. A request takes a single translation target language. Public copy describes roughly ten mixed languages for long-form transcription (including Chinese, English, Japanese and Korean) and mutual translation among Chinese / English / Japanese / Korean, with additional source languages into Chinese or English. Default RPM on the model card is 600.
Side-by-side facts (checked October 15, 2026)
| LecSync Realtime Speech API | Fun-ASR-Realtime | Paraformer-realtime-v2 | Gummy-realtime-v1 | |
|---|---|---|---|---|
| Role | STT + live MT + speakers + digest | Streaming STT | Streaming STT (legacy) | Streaming STT and/or MT |
| List price (Beijing) | $0.006/min ($0.36/hr), every feature | $0.000047/s ≈ $0.17/hr | CNY 0.00024/s ≈ $0.12/hr | CNY 0.00015/s ≈ $0.077/hr per meter |
| With live translation | Included | Not on this model — add MT | Not on this model — add MT | Second meter at the same rate ≈ $0.15/hr |
| Languages | 60, any→any, one pair per stream | Multilingual + Chinese dialects | ZH (incl. dialects) / EN / JA / KO (+ DE/FR/RU on some IDs) | ~10 for STT; ZH/EN/JA/KO mutual MT; others → ZH or EN |
| Speakers during the session | Yes, on every final segment | Not on realtime (overview) | Not on realtime (overview) | Not advertised as diarization |
| Live minutes / digest | digest: true on the same socket | Not offered | Not offered | Not offered |
| Custom terms | Glossary + forced translation pairs | Hot words | Hot words | Hot words |
| Default RPM (public) | Not published | 1,200 (Beijing + Singapore) | 1,200 (Beijing) | 600 (Beijing) |
| Regions / gateways | 6 gateway regions | Beijing; Singapore International WebSocket | Beijing | Beijing card only |
| Audio (typical) | 16 kHz mono PCM (v1) | pcm/wav/mp3/opus/…; any rate on Fun-ASR | Any sample rate | VAD-segmented stream |
| Vendor posture | Flat minute, best-effort | Current Fun-ASR line | "Migrate to Fun-ASR or Qwen-ASR" | Separate STT/MT meters |
What 1,000 hours cost
Alibaba rows use Beijing list prices. CNY-only cards are converted with the Fun-ASR dual listing (~7.02 CNY/USD). The Google Cloud Translation NMT stand-in ($20 per million characters after 500,000 free, assuming ~50,000 characters of transcript per hour) is only for the Fun-ASR + external MT path.
| 1,000 hours of live captions | LecSync | Alibaba Model Studio |
|---|---|---|
| Transcription only | $360 | Fun-ASR ≈ $169; Paraformer ≈ $123; Gummy STT ≈ $77 |
| + live translation | $360 | Gummy STT+MT ≈ $154; Fun-ASR + Google NMT ≈ $1,159 |
| + speaker labels during the session | $360 | Not on the realtime ASR cards above (use offline filetrans, or another vendor) |
| + live structured minutes | $360 | Not listed on Fun-ASR / Paraformer / Gummy realtime |
The second column is calculated from public rates, not a quote. Resource-plan discounts and console promotions can pull Alibaba STT even lower; those plans are out of scope here. For a wider vendor grid, see what 1,000 hours cost with translation, speakers and summaries.
Where Alibaba Model Studio is the better pick
- Chinese streaming STT at list price. Fun-ASR at about $0.17/hr and Paraformer at about $0.12/hr undercut LecSync's $0.36 whenever you only need the source-language transcript.
- Dialect and hot-word work inside China. Fun-ASR's overview stresses Mandarin dialects, noise robustness and large-scale hot-word customization — a natural fit for China-facing call centers and classrooms.
- Gummy when your language pair fits. If you live inside ZH/EN/JA/KO mutual translation (or the documented one-way pairs into Chinese or English), Gummy's combined meters can land near $0.15/hr — still below LecSync — without standing up a second vendor.
- Alibaba Cloud footprint. DashScope SDK, Beijing and Singapore WebSocket endpoints, and 1,200 RPM defaults on Fun-ASR make sense when the rest of the stack already sits on Aliyun.
- Offline diarization in the same family. When speaker labels can wait until after the session, Fun-ASR / Qwen file-transcription models advertise diarization on the overview table.
Where LecSync fits better
- Captions in a second language across 60 options. Translation streams with the transcript at about +88 ms and comes back paired by segment ID. More on the trade-off in built-in translation vs DIY STT + MT.
- Who said what, while it's being said. Every final segment can carry a speaker label with
startMs/endMs. See building a meeting client with speakers and timestamps. - Notes during the meeting.
digest: truepushes sections of minutes over the same socket. - One USD minute rate. Features on or off do not change $0.006/min; empty connections are not billed.
- Latency that doesn't drift. After a network stall, backlog is capped at 2 seconds. More in how LecSync caps backlog at 2 seconds.
- Gateways outside a single cloud region story. Six gateway regions, including nodes useful when speakers are not only in China.
FAQ
Does Fun-ASR-Realtime include translation?
No. Fun-ASR-Realtime is a speech-to-text model. For live translation on Alibaba Model Studio you look at Gummy-realtime-v1 (separate STT and MT meters) or newer LiveTranslate realtime models that bill in tokens. LecSync keeps transcription and translation on one WebSocket and one minute rate.
How is Gummy billed when both recognition and translation are on?
Public WebSocket docs say the two capabilities are billed separately at the same CNY-per-second rate. Turning both on for the same second of audio therefore doubles the Gummy meter relative to transcription alone.
Can Alibaba realtime ASR label speakers during the stream?
The Model Studio speech-to-text overview marks speaker diarization as unsupported on Fun-ASR-Realtime and the other realtime recognition rows shown there. Diarization is listed on offline / file-transcription models. LecSync returns speaker labels on final segments when speakers: true.
Is Paraformer still the right default?
Alibaba's overview labels Paraformer an older-generation family and recommends migrating to Fun-ASR or Qwen-ASR when possible. Paraformer-realtime-v2 remains on the price list (CNY 0.00024/s) and still covers Chinese dialects plus English, Japanese and Korean.
What does LecSync include in $0.006 per minute?
Realtime transcription in 60 languages, realtime translation between any two of them, audio-aligned timestamps, speaker labels, glossary and domain context, AI translation polish, live structured minutes, and optional 30-day storage of the recording and transcript. Turning features on or off does not change the rate.
Bottom line
Pick Alibaba Model Studio when the job is Chinese-heavy streaming transcription (or a Gummy language pair) inside Aliyun and list-price STT is the priority. Pick LecSync when the live product needs translation across a wide language matrix, speaker-aware captions and running minutes on one flat dollar-per-minute bill. Start with a short proof stream on the developers page — POST /v1/realtime/connect, then one WebSocket.