Baidu AI Cloud (百度智能云) sells two streaming speech products that matter for live captions. Real-time ASR (实时语音识别) turns a Mandarin or English audio stream into text over WebSocket. Real-time speech translation (实时语音翻译) is a separate WebSocket API that returns the transcript plus a translation in one of 45 languages. Speaker labels are not part of either.
The LecSync Realtime Speech API puts all of that on one connection: transcript, translation between any two of 60 languages, a speaker on every final segment and running minutes, for $0.006 per audio minute ($0.36 per hour).
Baidu figures below come from its official pages, read on October 20, 2026: the speech pricing doc, the real-time ASR WebSocket API, the speech product overview, the free test quota page, the real-time speech translation API and its pricing, plus text translation pricing. Prices are in yuan. Dollar figures use 6.72 CNY per USD, the CFETS central parity for October 20, 2026.
The short answer
For Chinese or English captions in one language, on a team that already runs on Baidu Cloud inside mainland China, Baidu real-time ASR is cheaper. A 1,000-hour package costs ¥1,800 (about $268), against $360 on LecSync, and larger packages go down to ¥0.9 per hour.
Once you need translated captions, Baidu's own real-time speech translation costs ¥10 per hour postpaid, or ¥9,000 (about $1,339) for a 1,000-hour package. LecSync is $360 for the same 1,000 hours, with speaker labels and live minutes included. LecSync also covers any pair across 60 languages and does not need a mainland China cloud account.
How each option works
LecSync. Your server calls POST https://api.lecsync.com/v1/realtime/connect with the audio format, the language pair and the features you want, then streams 16 kHz mono PCM to the returned URL. The same socket sends preview and final transcript segments, translations paired by segment ID, speaker labels with millisecond timestamps (speakers: true) and minutes section by section (digest: true). Language is detected automatically, and translation direction is picked per segment, so a Chinese and English conversation needs one stream. The first word reaches the screen about 0.28 seconds after speech starts, and translation adds roughly 88 ms.
Baidu real-time ASR. You open wss://vop.baidu.com/realtime_asr, send a START frame with your app ID, API key and a dev_pid model, then stream 16 kHz, 16-bit mono PCM in 20 to 200 ms frames (160 ms recommended). The public models are Mandarin (1537 light punctuation, 15372 full punctuation), a multi-dialect model (15376, covering Mandarin, Cantonese, Sichuanese and Northeastern Mandarin) and English (1737 without punctuation, 17372 with). You choose one model per stream. Each sentence comes back as MID_TEXT previews and a FIN_TEXT final with start_time and end_time in milliseconds.
A few limits shape the integration. One call can carry at most one hour of audio. A sentence longer than 60 seconds is cut. If the server receives no audio for 5 seconds, it returns an error and closes the connection. After a network drop, the docs ask you to cache audio on the client, open a new request from the last end_time and add that offset to the new timestamps yourself.
Baidu real-time speech translation. A separate endpoint, wss://aip.baidubce.com/ws/realtime_speech_trans, takes one from and one to language in its START message and accepts PCM at 8, 16 or 44.1 kHz. It returns interim and final results with the source sentence and its translation, and can optionally stream spoken translation as MP3. The documented result fields carry no timestamps and no speaker. The server closes the session after 30 seconds without a message.
Baidu also lists a "极速版" (fast edition), but it is short-audio recognition for files under 60 seconds, Mandarin only, billed per request. It does not stream.
Side-by-side facts (checked October 20, 2026)
| LecSync Realtime Speech API | Baidu real-time ASR | Baidu real-time speech translation | |
|---|---|---|---|
| Price | $0.006/min ($0.36/hr), all features | ¥3/hr postpaid (≈ $0.45); packages ¥1.8 to ¥0.9/hr | ¥10/hr postpaid (≈ $1.49); packages ¥9.5 to ¥8.5/hr |
| Languages | 60, auto-detected | Mandarin, English, plus a dialect model; one per stream | 45 |
| Live translation | Any two of 60, same socket | No | Yes, one fixed from/to per session |
| Speaker labels | Every final segment | Not documented | Not documented |
| Timestamps | startMs/endMs on every final, kept across reconnects | Per sentence, in ms; you re-offset after reconnecting | Not in the documented result fields |
| Live minutes | digest: true | No | No |
| Audio input | 16 kHz mono PCM16 | 16 kHz mono PCM16 only | PCM16 at 8, 16 or 44.1 kHz |
| Session limits | 2-second backlog cap | 1 hour of audio per call; 5 s silence-to-disconnect | 30 s idle timeout |
| Default concurrency | Not published | 50 paid; 3 or 5 on the free tier | 10 |
| Extra concurrency | — | ¥100 per stream per month, or ¥1,000 per year | Through sales |
| Free trial | — | 10 hours | 2 hours (personal) or 10 hours (business) |
| Account | Register and create an API key | Real-name verified Baidu Cloud account | Same |
What Baidu real-time ASR costs
| Plan | Price | Per hour | In USD (6.72) |
|---|---|---|---|
| Postpaid | ¥3/hr | ¥3.00 | $0.45 |
| 1,000-hour package | ¥1,800 | ¥1.80 | $0.27 |
| 10,000-hour package | ¥15,000 | ¥1.50 | $0.22 |
| 100,000-hour package | ¥120,000 | ¥1.20 | $0.18 |
| 500,000-hour package | ¥450,000 | ¥0.90 | $0.13 |
Packages are valid for one year, and usage falls back to postpaid once they run out. Billing is by call duration and settles hourly. Baidu's pricing doc says calls that fail because of authentication or audio quality are still counted. The product page also shows a new-customer promotion "from ¥0.2/hr" on a first purchase. Its terms are not in the pricing doc, so it is left out of the math here.
What 100 and 1,000 hours cost
| Monthly audio | LecSync | Baidu real-time ASR | Baidu real-time speech translation |
|---|---|---|---|
| 100 hours, same-language captions | $36 | ¥300 ≈ $44.64 postpaid | — |
| 1,000 hours, same-language captions | $360 | ¥1,800 ≈ $268 (package) | — |
| 100 hours, translated captions | $36 | — | ¥950 ≈ $141 (package) |
| 1,000 hours, translated captions | $360 | — | ¥9,000 ≈ $1,339 (package) |
| 1,000 hours, ASR + Baidu text translation (DIY, English speech) | $360 | ¥4,446 ≈ $662 (estimated) | — |
| 1,000 hours, ASR + Baidu text translation (DIY, Chinese speech) | $360 | ¥2,535 ≈ $377 (estimated) | — |
| Speaker labels or live minutes | Included | Not available | Not available |
The DIY rows send only final sentences to Baidu's general text translation at its postpaid rate of ¥49 per million characters, ignoring the one-time free allowance. They assume about 54,000 characters per hour of English speech and 15,000 per hour of Chinese, which are estimates. For a wider comparison across vendors, see what 1,000 hours cost once translation, speakers and summaries are added.
Access from outside mainland China
Baidu's speech APIs are documented and sold on cloud.baidu.com. Its billing overview requires personal or business real-name verification before paid use. On October 20, 2026, Baidu AI Cloud's international site (intl.cloud.baidu.com) featured compute, network and storage products, and its old speech product URL redirected to a 404. If your users and company sit outside China, check account eligibility with Baidu sales first.
LecSync needs only a registered account and an API key. It routes each stream to one of 6 gateway regions, or to the nearest healthy node, so audio does not have to cross an ocean.
When Baidu is the better pick
- Same-language Chinese captions at volume. From ¥1.8 down to ¥0.9 an hour, Baidu undercuts LecSync's $0.36 by a wide margin.
- Dialects. The
15376model covers Cantonese, Sichuanese and Northeastern Mandarin alongside standard Mandarin. - Custom language models. Baidu's self-training platform lets you upload vocabulary and long text to train a Mandarin model, then call it through
lm_id. - Spoken translation output. Real-time speech translation can return MP3 of the translated text, with a male or female voice option for English.
- Private deployment. Baidu offers on-premises speech packages for CPU and GPU servers, including domestic Chinese hardware.
- A stated uptime figure. The postpaid card lists 99.9% service stability. LecSync is best-effort today, with no SLA.
When LecSync fits better
- Translated captions on a budget. $0.36 an hour against ¥10 (about $1.49) postpaid on Baidu, with any pair from 60 languages. How built-in translation saves work is covered in built-in translation vs a DIY STT + MT pipeline.
- Two-way conversations. One LecSync stream translates Chinese into English and English into Chinese, choosing the direction per segment. Baidu fixes one
fromand onetoper translation session. - Knowing who said what. LecSync adds a speaker label and audio-aligned timestamps to every final segment. See building a meeting client with speakers and timestamps.
- Long lectures and meetings. Baidu ASR caps a call at one hour of audio and leaves reconnect stitching to you. LecSync keeps segment times on one recording clock across upstream reconnects and caps backlog at 2 seconds, as described in how LecSync caps backlog at 2 seconds.
- Minutes during the session.
digest: truepushes confirmed sections over the same socket. - Products outside China. No Chinese cloud account to set up, and 6 gateway regions to choose from.
If you are weighing other Chinese clouds too, see LecSync vs Tencent Cloud real-time ASR and LecSync vs iFlytek real-time ASR.
FAQ
How much does Baidu real-time speech recognition cost per hour?
As of October 20, 2026, Baidu AI Cloud charges ¥3 per hour postpaid for real-time ASR in Mandarin or English, about $0.45 at 6.72 CNY/USD. Prepaid packages bring it to ¥1.8 per hour at 1,000 hours and ¥0.9 per hour at 500,000 hours. Packages last one year. LecSync charges $0.36 per hour with translation and speakers included.
Does Baidu real-time ASR support speaker diarization?
Baidu's real-time ASR WebSocket documentation describes no speaker field as of October 20, 2026. Final results carry the sentence text with start and end times only, and the overview asks for audio without several people speaking at once. LecSync labels every final segment with a speaker when speakers: true is set.
Can Baidu translate live speech in real time?
Yes, through a separate real-time speech translation API that covers 45 languages. Each session sets one source and one target language. It costs ¥10 per hour postpaid, or ¥9,000 for a 1,000-hour package, with 10 concurrent streams by default. LecSync includes translation between any two of 60 languages at $0.006 per minute.
Is the Baidu speech API available outside China?
Baidu sells its speech APIs on cloud.baidu.com, and paid use requires a real-name verified account. On October 20, 2026, the international site listed no speech product. Teams outside mainland China should confirm eligibility with Baidu sales. LecSync needs only a registered account and an API key, and runs 6 gateway regions.
Bottom line
Choose Baidu real-time ASR when you need Mandarin or English captions in a single language, inside mainland China, at the lowest cost per hour, or when dialect models and custom training matter. Choose LecSync when captions have to be translated, labelled by speaker, longer than an hour, or used outside China. All of that runs on one socket at $0.006 per minute. You can open a test stream from the developers page with a single POST /v1/realtime/connect.