Model or dataset
wwbin2017/bailing avatar
wwbin2017/bailing

bailing: an 800ms claim with no measurement, and fallback tags that predate the feature they back up

Bailing is a GPT-4o-like voice chat robot built on ASR + LLM + TTS, integrating excellent models like DeepSeek R1 and connecting to OpenClaw. A true personal voice assistant with latency as low as 800ms, able to run on low-spec machines like Mac, and supports interruption.

1,795 stars307 forksPythonMIT

At a glance

What is it?
bailing is a Chinese language voice assistant built from FunASR, silero-vad, DeepSeek and four speech synthesis engines, wired to OpenClaw as its tool calling layer. The repository front page is written in Chinese, promises end to end latency of 800ms without a GPU, and its install steps depend on a directory the repository listing does not contain.
Who is it for?
bailing fits someone who wants a Chinese language voice assistant on their own machine, is comfortable pointing it at a DeepSeek key, and does not need it to work on a phone today. Read the requirements file before you plan hardware, because the 800ms figure and the no GPU claim are not backed by anything in the repository, and the file pulls two model runtimes, a transformer library, four speech synthesis paths and five audio input and output packages.
Can I use it commercially?
Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
Is it still maintained?
Activity is slowing. The repository last received commits 7 months ago.
What is it written in?
Mainly Python, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • wwbin2017/bailing1,795 ★ · +30
Star history
Repository2026-09-102026-11-15
wwbin2017/bailing1.8k1.8k

30 Stars in the last 66 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

The 800ms figure is the headline and nothing on the page measures it

The latency number is the project's main claim. It appears in the project description as latency as low as 800ms, in the introduction as an end to end latency of 800ms, and the first bullet under project features promises a low latency, stutter free conversation that feels close to talking to a person.

No page in the repository contains a measurement. There is no hardware description, no audio hardware, no model sizes, no network hop, no percentile, and no indication of whether 800ms is a floor, a median or a best case. The two statements also differ from each other: one says as low as, the other states the figure flatly.

The claim sits next to two other unbacked ones. Lightweight deployment is stated as needing no high end hardware and no GPU, runnable locally, while still delivering GPT-4 class performance, and the description adds that low configuration machines such as a Mac can run it. Against that, the requirements file is worth reading directly, because it is where the weight actually lands.

What the file pins is a full local stack: `funasr==1.1.6`, `silero_vad==5.1`, `transformers==4.41.1`, `kokoro>=0.8.1`, `misaki[zh]>=0.8.1`, `chattts==0.1.1`, plus `torch` and `torchaudio` with no version at all. Running four speech synthesis engines and two model runtimes locally on a machine with no GPU is achievable. Describing it as lightweight is a claim about the hardware bill, not about the dependency list.

AIGC arrived in v0.0.3 and the suggested fallbacks are v0.0.1 and v0.0.2

The release titles are short. v0.0.1 on 2024-10-04, v0.0.2 on 2025-03-14, and v0.0.3 on 2025-05-31, whose title reads that AIGC capability was added.

Now read the configuration step. To use the general AIGC configuration, which is described as still in testing, the page says that if it is unavailable you can use the tag branches v0.0.1 or v0.0.2.

Those are the two releases that predate the feature. AIGC capability entered in v0.0.3, so the two suggested fallbacks are the versions without it. The advice is stated plainly on the page and cannot be followed, which suggests the tags were written down before v0.0.3 was cut rather than checked afterwards.

The same step also asks for a second requirements file and a second configuration file inside a vendored dependency, and that is the general AIGC path. Its configuration is a TOML file needing `model`, `base_url` and `api_key`, separate from the YAML that configures ASR and the LLM. So the project has three configuration surfaces: `config/config.yaml` for the model pipeline, `config/.env` for the OpenClaw authorisation, and the TOML inside the vendored tree for AIGC.

The last recorded push is 2026-04-06, and the newest release predates it by over ten months, so the branch has moved since the last tag without a new tag to mark it.

Flask-SocketIO is required twice, under two names and two constraints

The requirements file is twenty seven lines, and one package appears twice:

bash
Flask_SocketIO==5.3.7
bash
Flask-SocketIO~=5.3.7

The two lines differ in the character joining the distribution name to its version constraint. The first spells the name with a low line and pins an exact release. The second uses a hyphen and a compatible release clause. Those are two different strings to a package manager, so the file is asking for two installs of what is one library, and it accepts a range in one place where it demands an exact pin in the other.

The wider file mixes four pinning styles. Most entries are exact, down to `PyYAML==6.0.2` and `pydub==0.25.1`. `kokoro` and `misaki` use floors at 0.8.1, and `misaki` also requests the `zh` extra. `Flask-SocketIO` uses a compatible release clause. And three entries carry no constraint at all.

Those three are the ones to notice. `torch`, `torchaudio` and `dotenv` are written with no version, so a fresh install resolves whatever those distributions currently point at, while `transformers` beside them is held at 4.41.1. A model runtime that floats against a transformer library that does not is the kind of pairing that breaks on a rebuild rather than on a change anyone made here.

The install needs a vendored directory that is not in the repository listing

Step two of the install runs two pip commands:

bash
pip install -r requirements.txt
pip install -r third_party/OpenManus/requirements.txt

Step three then asks you to configure `third_party/OpenManus/config/config.toml` with a model, a base URL and an API key.

The repository's top level has sixteen entries: `.gitignore`, `LICENSE`, `README.md`, `README_en.md`, `assets/`, `bailing/`, `config/`, `documents/`, `main.py`, `models/`, `plugins/`, `requirements.txt`, `server.py`, `server/`, `static/` and `tmp/`. There is no `third_party` among them.

So a fresh clone does not obviously contain the path that two of the documented install steps depend on. Either the directory arrives through something the page does not mention, or it is expected to be placed there by hand, or the AIGC path is reachable only from a checkout that differs from the default one. The page does not say, and the second pip command has no alternative form for someone who only wants the voice pipeline.

Two other manual steps sit in the same list. The ASR model has to be downloaded by hand from a Hugging Face repository into `models/SenseVoiceSmall`, so the repository ships no weights. And the LLM step asks for a DeepSeek API key from the provider's own console, while noting that OpenAI, Qwen, Gemini and one other provider can be configured instead. Only the OpenAI client is in the requirements file, so the alternatives are reached through a compatible base URL rather than through their own SDKs.

A self-signed certificate is generated, then the page sends you to http

The recommended path is the server run, and it starts by generating a certificate:

bash
openssl req -x509 -newkey rsa:4096 -keyout key.pem -out cert.pem -days 365 -nodes

RSA at 4096 bits, a 365 day lifetime, and `-nodes`, which writes the private key to disk with no passphrase. The section labels this a development environment certificate.

The next instruction is to start the service and then open `http://localhost:8000` and click the start button. Plain HTTP. The certificate that was just generated is not used by the URL the page tells you to visit, and nothing in the section says which of the two the service actually serves, or how the certificate should be presented in a browser that has never seen it.

The two run instructions also disagree about where the code lives. The local run does `cd server` before `python server.py`, and the server run says the same command works without the `cd`. The top level has both a `server.py` file and a `server/` directory, so the two instructions point at different relative paths from different working directories and the page does not say whether they are the same file.

A mobile browser is the reason the certificate exists, since a phone will refuse a self signed certificate for a plain HTTP origin. That case is not spelled out, and the certificate has no passphrase while being valid for a year, so it is the kind of file that gets committed to a repository by accident if the working directory is inside one.

The licence file permits commercial use and the disclaimer forbids it

The repository carries an MIT LICENSE file at the root, and the licence section states the project is released under the MIT licence with the original notice to be retained.

A disclaimer section says something different. The project is for personal study and research, is not suitable for commercial use or a production environment, use may cause data loss or system failure, and the project provides no technical support or guarantee of any kind, with the user bearing the risk.

Those two documents are in direct conflict, and the one that matters legally is the licence. MIT grants the right to use, copy, modify, merge, publish, distribute, sublicense and sell, and the only condition is retaining the notice. Nothing in that grant can be withdrawn by a paragraph in the README, so the disclaimer does not narrow what the licence permits. What it does do is tell a reader the author does not intend to support it and does not stand behind it, which is a different and useful thing to know.

The disclaimer also says that if you do not agree with the terms you should not use the project, which frames a personal preference as a condition of use. Treat the two documents as answering different questions: the licence answers what you are allowed to build, and the disclaimer answers who will answer the phone.

Given the same caution applies to the tools the assistant can reach. `open_application` will start a named application on a Mac, with the example of opening Safari, and `web_search` will issue a query you did not type.

The interruption table models the voice trigger and not the keyword one

Interruption is one of the headline features, described as a flexibly configurable policy that can recognise both keywords and voice interruption.

The state table has two axes, whether the player is playing and whether the user is speaking, and four rows:

| Player | Speaking | Behaviour | | --- | --- | --- | | playing | no | normal | | playing | yes | interruption | | not playing | no | normal | | not playing | yes | VAD decides, ASR recognises |

Every row describes the voice path. The VAD decides that speech is present and ASR turns it into text, and when the player is already running the same two conditions are labelled the interruption case. The keyword half of the policy has no row, no column and no state, even though it is named in the same feature description as the voice half.

So the table is a description of the audio path, not of the interruption policy. There is no row for a keyword arriving while the player is silent, no row for a keyword arriving during playback, and nothing indicating whether a keyword is checked before ASR runs or after it returns. A configuration described as flexible is documented as a two by two grid of the speech case only.

The remaining roadmap gives one more piece. Eight items are ticked, including RAG and agent work, memory, OpenClaw integration, and WebRTC and WebSocket support. One is not, and it is voice wake up.

Editorial conclusion

bailing fits someone who wants a Chinese language voice assistant on their own machine, is comfortable pointing it at a DeepSeek key, and does not need it to work on a phone today. Read the requirements file before you plan hardware, because the 800ms figure and the no GPU claim are not backed by anything in the repository, and the file pulls two model runtimes, a transformer library, four speech synthesis paths and five audio input and output packages. Install the main requirements and leave the vendored OpenManus step out unless you need the general AIGC path, and note that the directory it names is not in the repository listing. Do not rely on the documented fallback tags for AIGC, since both predate the feature. Decide how you will handle the certificate yourself, because the page generates one and then points you at plain HTTP, and the key it writes has no passphrase. Treat the MIT licence as the permission you actually have and the disclaimer as a statement about support, since the disclaimer cannot narrow the grant. If you are evaluating this for anything involving scheduling, opening local applications or web search, read the tool list rather than the feature bullets, and assume the interruption behaviour is the voice path shown in the table until you have tested the keyword path yourself.

Frequently asked questions

What models does bailing use for speech and language?

FunASR for automatic speech recognition, silero-vad for voice activity detection, DeepSeek as the language model, and edge-tts with Kokoro-82M, ChatTTS and the macOS say command for speech output. Configuration lives in config/config.yaml, and the ASR weights are downloaded separately into models/SenseVoiceSmall.

What does bailing need before it will run?

Python 3.12 or later and pip, then the dependencies in requirements.txt. A second file inside third_party/OpenManus is also installed by the documented steps, and the SenseVoiceSmall model has to be downloaded by hand. The repository listing does not include a third_party directory.

Can bailing interrupt itself while it is speaking?

That is the scenario the state table labels the interruption case, where the player is running and the user is speaking. The page also describes a configurable policy that recognises keywords as well as voice, but the table itself has no row for a keyword trigger in either player state.

What licence is bailing released under?

MIT, with a LICENSE file at the repository root and a licence section that says the original notice must be retained. A separate disclaimer states the project is for personal study and research and is not suitable for commercial use or production, which does not narrow what the MIT grant allows.

Which tools can bailing call?

Seven are listed: get_weather, ielts_speaking_practice, get_day_of_week, schedule_task, open_application, web_search and aigc. open_application starts a named application on a Mac, and aigc is described as a general purpose entry point connected to openclaw, with the page noting that its configuration is still in testing.

Official sources

  1. Issues
  2. License: MIT
  3. README
  4. Releases
  5. wwbin2017/bailing on GitHub
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/wwbin2017-bailing.svg)](https://hysenlabs.com/projects/wwbin2017-bailing)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/wwbin2017-bailing.svg)](https://hysenlabs.com/projects/wwbin2017-bailing)

Was this useful?