# One reply earned 300 stars: VoiceKey borrows speech recognition from seven IMEs onto one hotkey · GitHub open-source project

> Source: https://stringers.io/en/github/gh_J3n5en__voicekey.html (Stringers / 什么值得做, generated 2026-11-08)
> Stringers answers one question every day: what small product can one or two people build in a few months and actually make money from. Numbers come from 15 databases we collect daily, each with source and date.
> For agents: these verdicts and the underlying databases (verified revenue, App Store / Google Play charts, ad libraries, GitHub, Product Hunt…) are queryable live over MCP — tools today / hot / products / directions / direction / data. Setup: https://github.com/komako-workshop/stringers (the user logs in at stringers.io → "配置 agent" to get a personal URL). Site index: https://stringers.io/llms.txt

> A creator dropped a link in the replies to a 90k-view tweet titled "Get Doubao's offline model." The original author then retweeted it with a line saying "use his." A Telegram news channel reposted it too. It fits an old demand exactly: recognize the big tech companies' voice, but don't use their input method.

[GitHub open-source project](https://stringers.io/en/github/index.md) · written 11-6

J3n5en/voicekey · J3n5en

# One reply earned 300 stars: VoiceKey borrows speech recognition from seven IMEs onto one hotkey

**What it is** A voice input tool for desktop and phone. You don't have to switch your input method. Hold a hotkey and speak: recognition from Doubao, WeChat and five other IMEs outputs text at the same time, you pick one to insert. It also includes two offline engines.

**Why it took off** The author left a link under a post with 90k views that was about "getting Doubao's offline model." The original poster then reposted with a line saying "use his," and a Telegram news channel shared it too. It hit an old need: you want big-tech recognition without a big-tech input method.

**C** **What you can take from it** Don't copy the reverse-engineering of big-tech APIs. Learn two moves from it: drop a link to a working product under viral posts on the same topic, and run several engines at once with a candidate box where the user picks one. Apply that to a compliant tool built on a local model plus the user's own official API key.

C grade: this narrow slot is already taken by a deep-tech, finished product. The ceiling is a few hundred to a thousand stars in the Chinese circle, and it's built on someone else's interfaces, so it could be taken down at any time and the stars go to zero. What's worth learning is the tactics and the interaction, not following this path. Whispers in small circles

Data (updated daily) · Repo created 2026-10-01 · 531 stars · 7 days +418

GitHub · stars

Data 2026-10-16 → 2026-11-08 (24 days, 24 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

A developer with just over 200 followers replied under a post with 90k views, saying "it includes WeChat offline and Doubao offline." In two days the repo went from 120 stars to 475. The work is truly hardcore, but it stands on someone else's interfaces. If you want to build a project that gets stars by copying it, what you can learn is the tactics, not this path itself.

### A one-line reply under a post

On Nov 5 2026, around 4 p.m. Beijing time, developer @FradSer posted on X: "Thanks to the big reverse-engineering era! I've already got Doubao's voice input offline model, connected it to the online model calls, replicated the business logic one-to-one, and deployed it into the service." The post reached 11-06 with 92.9k views.

Three hours later, @J3n5enNg dropped a line in the replies: "github.com/J3n5en/voicekey has WeChat offline and Doubao offline," with a GitHub card below it. Three hours after that, the original poster added "use his, nothing to do with me 😑" with the same link. In the same reply thread, someone asked, "Isn't this a bit risky?"

The repo was created in Oct 1 and reached 4 stars by Nov, 120 total. It had gained only 17 in the week before. Trendshift's daily numbers: 11-04 +15,11-05 +226,11-06 +114. On GitHub, by 11-06 p.m. it was at 475 stars and 119 forks.

### What it actually does

In one line: you don't switch input methods. Hold one key and speak. Cloud recognition from seven providers, Doubao, WeChat, Qianwen, Baidu, Sogou, iFlytek and Chatterfly, outputs text at the same time, and you press a number key to pick one for insertion. There are also two fully offline engines: WeChat offline (macOS, Windows) and Doubao offline (Apple-silicon Macs only).

The hero image in the README is this candidate box. While you speak, results from several providers stream in line by line. After you release the key, each result is finalized in turn and tagged with its latency. Only an aggregator would look like this. A big company's own input method wouldn't.

The hard parts are underneath: the protocols for all seven providers were reverse-engineered. The WeChat offline model uses an int8 inference the author rewrote from scratch, with no dependence on the official program. Doubao offline loads the engine library from Doubao's Android app directly inside a Mac process. In one month it shipped about 30 versions, expanding from macOS-only to Windows, Android (doesn't replace the IME, hold the volume-down key to speak), and iOS (a Chinese keyboard with voice, installed from source). It's built with Rust + Tauri + Svelte, and the repo has no license.

### Why now

The demand was already there. In 2026, Doubao, WeChat and Qianwen were all competing for "AI voice input" as an entry point. Most media comparisons said Doubao and WeChat had the best Chinese recognition, and it was free. But the official products required you to switch your system input method to theirs. Posts like "Looking for a voice input app" and "Does Doubao's Mac version require switching IME?" kept showing up on linux.do. VoiceKey answers exactly that: you get big-tech recognition without a big-tech input method.

The timing practically delivered itself. In early November, "AI makes reverse engineering easy" was a hot topic in Chinese tech circles (36 氪 had a piece titled "An open-source era? AI reverse engineering makes all software free"). FradSer's post was about "getting Doubao's offline model," but his own service wasn't open source. People who wanted the same thing saw the link in the replies, and clicking it took them to a finished product. The original poster's "use his" effectively handed this wave of traffic to him.

The second wave came from Telegram news channels. The OKHK channel posted "VoiceKey - a cross-platform voice input tool that integrates multiple speech recognition channels," and another channel reposted it with "If you don't want to run a model locally, you can try this online one… it works by reverse-engineering the client." The +114 on 11-06 is probably from these channels (an inference: I didn't find the exact timing or view counts of the channel posts).

I haven't seen coverage on Product Hunt, HN or Reddit, or any media reporting, and it hasn't reached the top of GitHub Trending. So far the heat is confined to small Chinese developer circles.

### Many stars, few users

Downloads (GitHub Releases, 11-06) are much more sobering: the v1.9.0 released on 11-01 had its Mac installer downloaded 66 times, Windows 96 times, Android 96 times. The file the desktop app pulls when it checks for updates was fetched 829 times. My estimate is that only a few hundred machines are actually in use.

The forks are odd too: 119 forks against 475 stars, about a quarter, and concentrated in the two days of the surge. One explanation is people forking it to keep a copy in case the repo gets pulled (an inference).

### Who else is doing the same thing

The "hold to speak" format is already crowded:

| Project | Scale | How it differs from VoiceKey |
|---|---|---|
| cjpais/Handy, OpenWhispr, anomalyco/hex | 37.6k / 10.8k / 5,300 stars | General models like Whisper, with weaker Chinese than Doubao or WeChat |
| HaujetZhao/CapsWriter-Offline | 7,081 stars, since 2023 | Open-source local models, fully offline |
| joewongjc/type4me | 1,542 star | Adds LLM polishing, doesn't use reverse engineering |
| Paxxs/Doubao-ime-hammerspoon | 208 star | Still requires the official Doubao IME, just auto-switches for you |
| lilong7676/doubao-murmur | 142 star | Connects only to the Doubao web version |
| Typeless, Wispr Flow, Lightning Say | Paid subscription | Paid, with AI polishing as the selling point |

In the narrow slot of "split out big tech's free recognition, aggregate it, and include offline," I've only seen this one building a complete multi-platform product. Some people on linux.do have stitched it together with existing reverse-engineering libraries and complained about low recognition accuracy, which suggests this step isn't something just anyone can do well.

### Its problems

Its differentiator is "borrowing": free, no key, no IME switch, all built on ByteDance's and Tencent's interfaces and models. The offline models are downloaded straight from their CDN. If an interface changes or gets blocked, retention depends entirely on the author chasing fixes. The reply thread's "Isn't this a bit risky?" isn't a joke: the most common fate for repos like this is a takedown notice, and the stars go with it.

### If it were you

**C. Learn from it, but don't build on this path.** Two things are worth taking: put a "works right now" finished product under viral posts on the same topic, and use an interaction where several engines run at once and the user picks from a candidate box. As a path, reverse-engineering big-tech APIs is already occupied by a deep-tech project. The ceiling is a few hundred to a thousand stars in the Chinese circle, and it could be taken down along with the repo at any time.

- **Who it suits** : people who already have some presence among Chinese developers on X and linux.do, and who can ship a complete product rather than just a protocol library. If what you want is a following and a reputation, and you don't care how long the repo survives.

- **Who it doesn't suit** : anyone who wants a project that accumulates stars over time and looks good on a resume. A takedown notice resets it to zero. Anyone who wants stars in the English-speaking world: nobody there is missing "free Doubao," and Handy, the general-model version, already has 37k stars.

- **How to build it** : move the aggregation interaction to a compliant base. For example, a hold-to-speak tool that runs local Whisper / SenseVoice alongside official APIs (Volcano, Tencent Cloud, OpenAI) using the user's own keys, with a candidate box to pick from. It has the same visible selling point, without the legal tail. Or apply "several providers return results at once, pick one" to translation or OCR, where the providers are roughly equal in quality and users can't tell which to choose.

- **How to get stars** : learn its tactics. Watch the X posts on your topic that are gaining traction, leave a line, a link and a clear image in the replies; put the screen that best shows the difference as the README hero image; and within two or three days of launch, ship several versions in a row so people who reposted see the project is alive.

- **When the window closes** : the heat around "AI reverse engineering" in Chinese circles probably has a few weeks left (estimate). If you want to ride this wave, have a finished product out before the end of November. Anyone building a compliant version isn't racing this window.

- **What would show you read it wrong** : ① VoiceKey is still around a month later with stars past 2,000, and ByteDance and Tencent have taken no action. That would mean the risk of this path is smaller than it looks, and aggregating reverse-engineered services is worth doing. ② A similar project appears in the English-speaking world (say, aggregating free translation or OCR services through reverse engineering) and gets several thousand stars. That would mean this tactic can break out. ③ The reverse: VoiceKey or a similar repo gets a takedown notice or its interfaces get blocked. Then even learning the tactics means switching to a different topic.

[GitHub repo ↗](https://github.com/J3n5en/voicekey)

**Research notes: competitor table, acquisition evidence, full research report**

Competitors

| Name | Where | Launched | Scale | How it differs from this one |
|---|---|---|---|---|
| Doubao IME for Mac | Big-company features | 2026-05 | ByteDance official, free | The official one requires switching the system input method. voicekey uses a global hotkey, doesn't require switching IMEs, and lets you use several providers together |
| WeChat IME for Mac 3.0 | Big-company features | 2026-10-15 | Tencent official, free | Same as above. voicekey also pulls out its offline package and runs inference itself |
| Typeless / Wispr Flow / Lightning Say | Other | 2025-2026 | Paid subscription, frequent in reviews | They charge and focus on AI polishing. voicekey is free and does recognition only |
| lilong7676/doubao-murmur | Open source | 2026-03-31 | 142 star | Only one channel, Doubao Web, with no offline mode |
| Paxxs/Doubao-ime-hammerspoon | Open source | 2026-03-31 | 208 star | Still relies on the official Doubao IME, just auto-switches |
| joewongjc/type4me | Open source | 2026-03-22 | 1542 star | Adds LLM polishing, doesn't use the free reverse-engineered channels |
| HaujetZhao/CapsWriter-Offline | Open source | 2023-05 | 7081 star | Uses open-source local models. voicekey uses big-tech models |
| cjpais/Handy, anomalyco/hex, OpenWhispr | Open source | 2025-2026 | 37.6K / 5.3K / 10.8K star | General models like Whisper and Parakeet, with Chinese recognition weaker than Doubao and WeChat |

Where users come from

- **Verified** **X: replies with links under a viral post** : @FradSer posted "Thanks to the great reverse-engineering era! Already pulled Doubao's voice input offline model" at 11-05, 92.9K views; @J3n5enNg posted the repo link in the reply section; at 14:25 UTC FradSer posted "Use his, nothing to do with me 😑" with the link, 1,637 views

- **Verified** **Telegram news channels** : OKHK channel post 11984 "VoiceKey - a cross-platform voice input tool integrating multiple speech recognition channels"; the Uncle Du netdisk channel reposted it, saying it "is implemented by reverse-engineering the clients"

- **Verified** **"AI reverse-engineering" topic heat** : 36 Kr's "Open-source era? AI reverse engineering makes all software free", plus Zhihu and Kuai Keji articles on REA

- Inferred **Author's own post (the mid-October small spike)** : 10-15, 10-16 gained 21 stars; the source was not found

## J3n5en/voicekey: why it took off

> In one line: an open-source desktop voice input tool that freely taps the speech recognition APIs of big Chinese tech companies' free input methods. It wires seven cloud channels (Doubao, WeChat, Qwen, Baidu, Sogou, iFlytek, Chatterfly) plus two offline models (WeChat offline, Doubao offline) into one global hotkey, on macOS, Windows, Android and iOS. Verdict: the product is genuinely hardcore, but this burst rode a piggyback on a hot topic: the author replied with the link under a 2026-11-05 post with 90K+ views about the "great reverse-engineering era", the original poster then passed the traffic along saying "use his", and Telegram news channels reposted it. The heat so far is limited to a small circle of Chinese developers, and the copyright and legal risk is high.

### 1. Basic data

- Repo created 2026-10-01, Rust + Tauri + Svelte (the old Swift implementation is kept on the legacy branch), no license (GitHub API, 2026-11-06).

- Stars: the data card recorded 11-05 at 135 and 11-06 at 344 (+209 in one day; only +7 in the first 17 days); the GitHub API already showed 475 around 14:20 UTC on 11-06, with 119 forks.

- Trendshift daily data: 11-04 +15, **11-05 +226** , 11-06 +114; 10-15 and 10-16 had small spikes of +7 and +14 ([trendshift.io/repositories/306361](https://trendshift.io/repositories/306361）)). Fork creation dates also cluster on 11-05 (59 forks) and 11-06 (29 forks) (GitHub forks API).

- Real usage is small (GitHub Releases download counts, 2026-11-06): the latest.json of v1.9.0 (11-01 release) was fetched 829 times, roughly equal to the number of in-app update checks on the desktop app; dmg downloads 66, Windows exe 96, Android apk 96; the older v1.6.1 latest.json was fetched 661 times. Estimated active installs are only a few hundred.

- Release cadence is fast: 10-01 to 11-01 saw about 30 releases in total, with platforms expanding from macOS only to Windows, Android and iOS (iOS only offers an unsigned ipa).

- Author J3n5en (X account @J3n5enNg) has 232 GitHub followers. Their EnsoAI has 998 stars, bye 1041 stars, EnsoCode 161 stars; they have some presence in the Chinese AI developer circle but aren't a big name.

### 2. Competitors

| Product | Where | Launched | Scale | How it differs from voicekey |
|---|---|---|---|---|
| Doubao Input Method for Mac (ByteDance) | Big-company features | 2026-05 (Tencent News 20260513) | ByteDance official, free | You must switch the system input method to Doubao to use it; voicekey needs no input method switch, works from a global hotkey, and can use several providers at once |
| WeChat Input Method for Mac 3.0 (Tencent) | Big-company features | 2026-10-15 (z.weixin.qq.com) | Official, free | Same as above, you have to switch the input method; voicekey extracts its online API and offline packages for separate use |
| Qwen PC voice input, Tencent Chatterfly, Zhipu AI input method | Big-company features | Launched within 2026 years | Official, free | Single vendor, tied to each one's client; voicekey aggregates these providers together |
| Typeless, Wispr Flow, Lightning Say | Others (paid apps) | 2025–2026 | Paid subscription, a regular in media side-by-side reviews | Paid, mainly marketed on AI polishing; voicekey is free, does no polishing, and only handles recognition |
| lilong7676/doubao-murmur | Open source | 2026-03-31 | 142 star | Only connects to one channel, the Doubao web version; voicekey has multiple channels and offline |
| Paxxs/Doubao-ime-hammerspoon | Open source | 2026-03-31 | 208 star | Uses Hammerspoon to auto-switch the input method and summon Doubao, still relying on the official input method; voicekey needs no official program at all |
| starccy, moziarnj07-sys's doubaoime-asr | Open source (library) | 2026 | A reverse-engineering library often cited in linux.do posts | It's only a protocol library; and linux.do users have reported that the recognition accuracy from the reverse-engineered output is low ([linux.do/t/topic/2196046](https://linux.do/t/topic/2196046）)). voicekey is a finished product |
| joewongjc/type4me | Open source | 2026-03-22 | 1542 star | macOS/iOS, with large-model polishing; does not take the free reverse-engineered channels |
| HaujetZhao/CapsWriter-Offline | Open source | 2023-05 | 7081 star | Local open-source model; voicekey uses the big companies' offline models |
| cjpais/Handy, anomalyco/hex, OpenWhispr, Open-Less/openless | Open source | 2025–2026 | 37.6K / 5.3K / 10.8K / 4.1K stars (our GitHub repos) | All are general models like Whisper or Parakeet; their Chinese recognition is not as good as Doubao or WeChat |
| BuildWithAIs/voicekey (same name) | Open source | 2026-01-12 | 184 star | Local recognition plus built-in LLM polishing, unrelated to this product, only the same name |
| FradSer's self-hosted Doubao offline service | Other | 2026-11-05 | Not open-sourced, only a single tweet | Takes the same reverse-engineering approach, only for personal use |

Verdict: the "hold to talk" form factor is already a red ocean, with many open-source and paid products. voicekey's niche is narrow: it reverse-engineers the recognition capability of big companies' free input methods and aggregates it, with no API key, no input method switch, and offline support. Within this subcategory, the doubao-murmur and hammerspoon approaches only connect one provider, or still depend on the official program.

### 3. Growth channels (where users come from)

1. **Replying with the link under a viral tweet (confirmed, the main cause)** . @FradSer posted at 2026-11-05 08:22 UTC: "Thanks to the great reverse-engineering era! Already pulled Doubao's voice input offline model, connected the online model calls, replicated the business logic one-to-one, and deployed it to the service", with 92.9K views ([x.com/FradSer/status/2118257675779850453](https://x.com/FradSer/status/2118257675779850453）)). The author @J3n5enNg posted in the reply section "github.com/J3n5en/voicekey has WeChat offline and Doubao offline", with a GitHub card. At 14:25 UTC, FradSer posted another "Use his, nothing to do with me 😑" with the repo link attached (1,637 views, [x.com/FradSer/status/2118348866294878382](https://x.com/FradSer/status/2118348866294878382），等于把想要他那套东西的人直接导到)), effectively routing everyone who wanted that stuff straight to voicekey. The star surge began right at 11-05 (+226).

2. **Repost by Telegram news channels (confirmed, timing inferred)** . The OKHK channel posted "VoiceKey - a cross-platform voice input tool integrating multiple speech recognition channels" ([tg.okhk.net/posts/11984, post number right next to](https://tg.okhk.net/posts/11984，帖子编号紧挨着) 11-06's post 12000; inferred to have gone out on 11-05 or 11-06). The "Uncle Du netdisk" channel also reposted it: "If you don't want to run models locally, try this online one. It has many different voice input sources built in, implemented by reverse-engineering the clients" (Jina search result summary). Channels like these drove the second wave (+114) of 11-06.

3. **The "big reverse-engineering era" topic itself (confirmed as background)** . 36kr's "Big Open-Source Era? AI Reverse Engineering Makes All Software Free" ([m.36kr.com/p/4027900020920708](https://m.36kr.com/p/4027900020920708）)), the REA article on Zhihu, and Kuaikeji's "the piracy era is back" all show that AI-assisted reverse engineering was a hot topic in Chinese tech circles around 2026 November. voicekey is a ready-made example of that topic.

4. The source of the mid-October small spike (10-15, 10-16 gained 21 stars) was not found; the author may have posted it on X or linux.do (unverified).

5. No evidence found of ads, Product Hunt, HN, Reddit, media coverage, or GitHub Trending placement (Trendshift shows it never made the top 25; our local x-radar corpus has no hits for voicekey).

### 4. Why it charted

- **The demand is real and already validated** : in 2026, Doubao, WeChat, Qwen and Sogou were all competing for the "AI voice input" entry point, media side-by-side reviews consistently rank Doubao and WeChat's Chinese recognition highest, and both are free. But the official products require switching the system input method, and posts like "looking for a voice input software recommendation" and "must I switch input methods to use Doubao on Mac?" keep recurring on linux.do. voicekey solves exactly "I want Doubao's or WeChat's recognition, but I don't want to switch input methods".

- **The technical barrier is real** : they rewrote the int8 inference of the WeChat offline model themselves (40 layers of Transformer, NEON/AVX2), loaded Doubao's Android ELF engine library directly inside a macOS process, and reverse-engineered all seven cloud protocols (protobuf/WebSocket, ECDH+AES, HMAC). On linux.do, someone used an existing reverse-engineering library and complained about low accuracy. The only complete, multi-platform finished product I've seen so far is this one.

- **The multi-channel candidate box is a visible selling point** : the README's first image is seven channels outputting text at once, with a number key to pick one to insert. This form only makes sense with aggregation, and an official product would never build it.

- **Timing** : 11-05's viral "great reverse-engineering" post was discussing the very same thing (using Doubao's offline model), the author replied in the comments, and the original poster handed the traffic over to him.

### 5. Differentiation

The differentiation is "free, no key, no input method switch, multi-provider aggregation plus offline", which in essence treats the ASR of big companies' free input methods as a public resource. That is indeed why it got reposted: both the tweet and the Telegram channels' repost copy emphasized "reverse engineering", "many voice input sources built in", "WeChat offline, Doubao offline", and nobody was talking about the experience. But this differentiation is built on other people's interfaces and does not constitute a moat:

- Legal risk: someone already asked under the original tweet "isn't this a bit criminal?". The repo has no LICENSE, and the offline models are downloaded straight from ByteDance's and Tencent's CDNs. The fork ratio is high (119/475, about 25%), and concentrated on the day of the burst, which looks more like forking to archive the repo in case it gets taken down (inferred).

- The interfaces could be changed or blocked at any time, and retention depends on the author keeping up with maintenance.

- It cannot charge money: the moment it does, it goes from "reverse-engineering sharing" to "commercial infringement".

### 6. Visibility and copycats

- Visibility: **it's circulating in small circles** . Reach is Chinese developer X (replies under a 90K-view tweet) plus two or three Telegram news channels; there's no media coverage and no revenue (the project is free, with no paid entry point).

- Copycats: similar ideas include FradSer's self-hosted Doubao offline service (not open-sourced), plus earlier doubao-murmur and hammerspoon approaches. After the burst a batch of forks appeared (e.g. tod-zhang/voicekey), but no one has yet been seen packaging one into a new product and releasing it.

### 7. Takeaways

- Dropping a link to a "ready-to-use finished product" in the reply section of a viral tweet on the same topic has a very high conversion rate. If the original author is willing to say "use his", the effect is equivalent to a free recommendation.

- The aggregation interaction of "multiple engines running at once + pick one from the candidate box" applies to any scenario where several models have similar capability and users can't tell which to choose, such as translation, OCR, or ASR.

- The real demand for Chinese voice input is "big-company recognition quality + no input method switch + a global hotkey". The proper route is to use the official ASR APIs from Volcano Engine or Tencent Cloud, billed per usage, then add a subscription, which would sidestep the legal risk of reverse engineering.

### 8. Uncertain

- The source of the October mini-wave.

- Exact posting times and view counts for the two channel posts from OKHK and 杜叔叔网盘.

- Retention: can only be roughly estimated from latest.json pull counts; can't be deduplicated by person.

- Whether ByteDance or Tencent will send DMCA notices or block the API.

This article was written by AI from public data and the data we collect daily. Numbers are sourced, and inferences are labeled.