# A day after a new model dropped, the only tutorial left that runs on Windows is for speech.cpp. How it got 358 stars in Japan by being fast to ship · GitHub open source

> Source: https://stringers.io/en/github/gh_nyosegawa__speech.cpp.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

> The author has 10k followers in Japan's AI community and posts a tweet with a video for every release. The official Python version of the new Japanese model was out, and it was first to ship a version that runs directly. When Haas-ja went viral, Japanese tutorials pointed to it.

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

nyosegawa/speech.cpp · nyosegawa

# A day after a new model dropped, the only tutorial left that runs on Windows is for speech.cpp. How it got 358 stars in Japan by being fast to ship

**What it is** A C++ voice tool you install on your own computer. One command installs it, and it can read text aloud, turn speech into text, and transcribe in real time. No Python, no CUDA. Works on Mac, Windows, and Linux, with GPU acceleration.

**Why it took off** The author has 10k followers in Japan's AI scene and posts a video tweet with every release. When the Japanese model came out, the official build was Python-only; he shipped a version that runs directly within hours. When Haas-ja went viral, Japanese tutorials pointed to his project.

**B** **How you could follow it** Stop making a Japanese version and do the same thing for another language (Korean, Indonesian, Arabic, etc.): within days of a new model's release, ship a one-line install with no Python, post a video tweet for each new model added, and get the people who write tutorials to know about it.

Grade B: the playbook is clear and can be copied for other languages, and the MIT license lets you fork it directly. But it has only reached a few hundred stars, mostly within Japan's circles. Its traction came from the author's existing audience and from shipping in the first days after a model drops. The field is crowded, and once a big project integrates a model, the window closes. What people in small circles are saying

Data (updated daily) · Repo created 2026-09-29 · 371 stars · 7 days +92

GitHub · stars

Data 2026-10-11 → 2026-11-08 (29 days, 28 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

In 2026 Oct 30, Japanese company Third Intelligence released the Japanese speech recognition model Haas-ja-v1.0, but the official code only ran on Linux with NVIDIA GPUs. The next day, Japanese developers wanted to try it on their Windows PCs, and nearly every tutorial they found used the same thing: speech.cpp, a C++ repo created only a month earlier. It had sat at about two hundred stars until then and gained 142 more in the following week. This piece covers what it is, why it was the one that won, and whether its approach can be learned by others.

### An RTX 3050 PC, 6 hours of audio in 2 min 45 sec

On the evening of Oct 31, tasekino on Qiita posted a review: using an RTX 3050 with 6GB of VRAM, he ran a 6 hour 10 minute lecture recording through Haas-ja for transcription. It took 2 min 45 sec, roughly 134x real time, and a short sentence came back in 0.03 to 0.04 seconds. He normally uses the cloud service Soniox, billed by the hour. The first conclusion of the article was:

> Windows でもすぐ動く。C++ 実装の speech.cpp に Windows / Vulkan 版のビルド済みバイナリがあり……WSL2 も CUDA も不要。 (Windows 上也能马上跑。C++ 实现的 speech.cpp 有 Windows / Vulkan 的现成程序……不用 WSL2,也不用 CUDA。)

He also wrote: the official inference code only supports Linux + NVIDIA + NeMo, so "on Windows, use speech.cpp."

At the same time, ai-hardware-zukan, Zenn, and note all had articles on "running Haas-ja on Windows / Mac," all using speech.cpp with model files the author had converted himself. The new model's buzz flowed through these tutorials straight to this repo.

### What it is

Put simply, it's a "voice toolkit" that lives on your own computer: one command installs it, and a command called `speech` can read text aloud (TTS) and turn audio into text (ASR). It also includes voice activity detection, speaker diarization, and noise reduction.

- Models are specified by name and downloaded automatically from Hugging Face on first use. The model files are GGUF conversions the author made himself and hosted under his HF account.

- Mac uses Metal, Windows and Linux use Vulkan, and without a GPU it falls back to CPU. No Python needed, no CUDA needed.

- `speech serve --open` opens a web page where you can try speech synthesis and transcription, including live transcription from your microphone. The same service exposes an OpenAI-compatible API, so a program that calls OpenAI's speech endpoints can switch to a local model just by changing the address.

Half the supported models are Japanese-specific: Irodori-TTS (Japanese synthesis), ReazonSpeech, parakeet-ja, and Haas-ja (Japanese recognition). The other half are Qwen3-TTS / Qwen3-ASR and NVIDIA's Nemotron family. The author's own positioning is clear: it is built for voice conversations and voice assistants, so "how fast the first sound comes out" is the main selling point. He measured Qwen3-TTS 0.6B at 0.029 seconds to first audio on an M5 Mac. On Zenn, someone tested Irodori-TTS on an M4 Pro: the official implementation had a median first-audio time of 1.50 seconds, speech.cpp 0.21 seconds, a difference of 7x.

### Who makes it

The author is Gyakuse (@gyakuse on X, about 10k followers), a well-known individual developer in Japan's AI scene whose earlier open-source projects had only a few dozen stars. His approach is consistent: with each major release, he posts a tweet with a demo video.

| Date | What happened | star |
|---|---|---|
| 09-29 | Created the repo, shipped v0.1.0 the same day | 0 |
| 10-09 | Published a bilingual Japanese-English long post on his blog, "I made speech.cpp" | — |
| 10-14 | Tweet: "for AI assistants, fast first audio, lightweight install, no Python/CUDA," 33.8k views | — |
| 10-15 | Hatena Bookmark (Japan's tech bookmarking site): 32 people saved it | 10-16 for 129 |
| 10-17 | Single day +32 | 10-18 for 180 |
| 10-19 to 10-30 | No new topic, 0 to 5 per day | Stalled at 216 |
| 10-31 | Tweet: new support for Nemotron speaker diarization, Nemotron-3.5-Transcribe, and Haas-ja-v1.0, with a 1 minute video; 43.8k views, 627 likes, 585 bookmarks | 219 |
| 11-01 | +60 in a single day, the highest of the month; Haas-ja tutorials appeared at the same time | 279 |
| 11-06 | +142 over the last 7 days, versus only +21 in the first 7 days | 358 |

(Star counts come from daily GitHub snapshots; view counts come from the X page, 11-06.)

Bookmarks and likes were about equal, suggesting many people were saving it to install later. Replies to the Oct 31 tweet included English- and Arabic-speaking users, but overall it never reached GitHub's trending page, had no dedicated threads on Hacker News or Reddit, and wasn't covered by major English-language accounts or outlets. The growth came almost entirely from Japanese X and Japanese tech communities (Hatena, Zenn, Qiita, note). Geographic data on stars isn't available, so "mostly Japanese users" is inferred from the channels.

### Why this one

1. **The new Japanese models only had Python versions, and it filled that gap.** The official implementations of Irodori-TTS, Haas-ja, and similar Japanese models are all PyTorch, so Windows users had to install WSL, CUDA, and a pile of dependencies. speech.cpp installs with one command and pulls models by name.

2. **Speed to ship.** Haas-ja was released on Oct 30, and support for its v0.9.3 was available the next day, with the model files converted and uploaded by the author. Tutorial writers couldn't find a second option, so they used it.

3. **Releases are the content.** In 38 days he shipped 29 versions, and nearly every version added a new model, each of which could be a tweet. His 10k followers gave him 2 to 40k views per post.

4. **The selling point is a number you can test.** "First audio in 0.029 seconds" and "7x faster than the official build" give people something to measure when they write comparison posts, and the results they get back vouch for it.

### There are many similar tools, and it only wins on the Japanese side

Local speech engines are a crowded space in 2026:

| Who | Scale | Compared with alternatives |
|---|---|---|
| whisper.cpp | 55k stars | Only Whisper recognition, no synthesis |
| sherpa-onnx | 16k stars | Covers phones and embedded devices, with more models, but slow to add new Japanese models |
| transcribe.cpp | 2,139 stars, 773 points on Hacker News (July) | Recognition only, 16+ model families, no Japanese specialization |
| CrispASR | 798 star | 24 recognition + 5 synthesis models, broader than it, but doesn't focus on Japanese or first-audio latency |
| NVIDIA NeMo-Speech.cpp | 242 star | Official NVIDIA project, runs only its own Nemotron models |
| mlx-audio | 8.2K star | Runs only on Apple chips |
| Speaches / Kokoro-FastAPI / LocalAI | 3.8K / 5.6K / 50k stars | Also offers an OpenAI-compatible API, but requires Python or Docker |

(Star counts from GitHub, 11-06.)

It can't match CrispASR on model count or transcribe.cpp on name recognition. Its only niche is packaging new Japanese models alongside Qwen3 and Nemotron in a program that needs no Python, with new models available the moment they drop. In the English community that difference isn't big enough, so its growth stalled at a few hundred stars within Japan's circles.

I haven't seen any repo copying it yet. On the other hand, someone is already using it as a base: the install instructions for the Japanese local voice-input tool OpenHinoVoice (51 stars) start with installing speech.cpp.

### What to learn from it

- **Watch for new open-source models in one language, and on the day of release, ship a version that runs with one command.** The people who write tutorials will distribute it for you, and the model's buzz flows to your project.

- **Convert the models yourself and host them, so the download is by name.** Removing the "set up the environment" step is what gets people to actually try it.

- **Add a new model with each version, and pair it with a video tweet.** The release itself is the content, so you don't need to come up with another topic.

- **Give a number people can reproduce.** First-audio latency or a speed multiple: once others measure it and write about it, you gain another entry point.

### If it were you

**Verdict** : its approach is worth learning from, but don't build another Japanese version. It already holds the Japanese niche, and its star count suggests the ceiling for a single-language project is a few hundred stars.

**Who it suits** : developers who can write C++, read ggml / GGUF, and port a new model within a day or two of its release. Ideally you already have some presence in a language's developer community (X, a tech blog, or a local community all work).

**Who it doesn't suit** : people who want to make money from it (this kind of project yields stars and reputation). People with no channel to speak through, since without an audience like the author's 10k followers and his 2 to 40k views per post, nothing will reach anyone. People who can't keep pace with model releases, since a week late and the tutorials are already written.

**How to do it** : pick a different language. Korean, Indonesian, Vietnamese, Arabic, and Hindi all have local teams releasing their own speech models, and the official releases are probably Python-only too (which ones lack a C++ port is unverified). Choose a language you know and that has a community, and build its recognition and synthesis models into a program that needs no Python and runs on both Mac and Windows. You could also fork speech.cpp (MIT license) and add models for a new language, skipping the framework work.

**How to get users** : each time you add a new model, post a video tweet where local developers hang out most (X, local tech blog platforms, Discord). Reach out to the people who write tutorials and tell them "here's how to run this on Windows." Leave a usage tip in the comments section of the model's Hugging Face page.

**When the window closes** : each time, it's "the first few days after a new model drops." The overall window is also narrowing: llama.cpp's main line has already started supporting Qwen3-TTS, and big projects like whisper.cpp and sherpa-onnx will eventually integrate popular models, at which point "only I can run this" stops being true (inference).

**What it means if you got it wrong** : the new model shipped a week ago and tutorials use a different tool, or the model's own team released a Windows build; or your launch post didn't get past 1,000 views; or developers in that language already rely on cloud services and nobody wants to run it locally.

[GitHub repo ↗](https://github.com/nyosegawa/speech.cpp)

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

Competitors

| Name | Where | Launched | Scale | How it differs from this one |
|---|---|---|---|---|
| whisper.cpp | Open source | 2022-09 | 55K stars | Only Whisper recognition, no synthesis and no new Japanese models |
| sherpa-onnx | Open source | 2022-09 | 16K stars | The ONNX route covers more ground, but it's slower to pick up new Japanese models and heavier to install |
| transcribe.cpp(handy-computer) | Open source | 2026-04 | 2,139 stars, HN 773 points | Recognition only: no TTS, no Japanese specialization |
| CrispASR | Open source | 2026-03 | 798 stars | A speech model hub with more models, but it doesn't focus on Japanese or first-audio latency |
| NVIDIA NeMo-Speech.cpp | Big-company features | 2026-07 | 242 stars | Only runs the Nemotron family; speech.cpp also wraps Nemotron and adds Qwen and Japanese models |
| qwentts.cpp / qwen3-tts.cpp / llama.cpp mainline Qwen3-TTS | Open source | 2026-02~05 | each 200-250-star | A single-model TTS port, no recognition or VAD |
| mlx-audio | Open source | 2024-11 | 8.2K stars | Apple MLX only, doesn't run on Windows/Linux |
| Irodori-TTS official | Open source | 2026-02 | 1,547 stars | PyTorch reference implementation, M4 Pro first-audio 1.50s; speech.cpp 0.21s (measured on Zenn) |
| Speaches / Kokoro-FastAPI / LocalAI | Open source | 2023-2024 | 3.8K / 5.6K / 50K stars | Also offers an OpenAI-compatible speech API, but needs Python/Docker and uses different models |

Where users come from

- **Verified** **The author's own X account (@gyakuse, 10k followers) posts with every release** : 10-14 tweet 33.8K views, about 415 likes; 10-31 tweet 43.8K views; 11-01 pinned tweet 21K views

- **Verified** **Hatena Bookmark** : 32 users, clustered in 10-15, matching the star count rising from 10-16 to 10-18 over 129–180

- **Verified** **Traffic from the new model launch (Haas-ja-v1.0) tutorials** : multiple Qiita, ai-hardware-zukan, Zenn and note posts on "running Haas-ja on Windows" all use speech.cpp + GGUF, published between 10-31 and 11-03, in the same window as a single-day +11-01 60

- **Verified** **Downstream project citing it** : OpenHinoVoice (51 stars) has the speech.cpp install command as its first install step; the author's own ASIST also uses it

- **Verified** **Author's blog and HF model repo** : a long bilingual Japanese/English post on nyosegawa.com; the HF sakasegawa/* GGUF models download automatically by name

- Inferred **HN / Reddit / GitHub trending** : the hackernews table has no posts; trending_days_any=0; no dedicated Reddit thread found

## Why nyosegawa/speech.cpp actually runs

> Channel: GitHub | repo created 2026-09-29 | about 358–366 stars and 22 forks as of 2026-11-06 | MIT | author Nyosegawa (@gyakuse on X, nyosegawa on GitHub, sakasegawa on HF)

### One line

A ggml-based C++ local speech toolkit: one `speech` command plus one `libspeech` library, running TTS (Qwen3-TTS, Irodori-TTS, Drama 3), ASR (Qwen3-ASR, ReazonSpeech, parakeet, Haas-ja, Nemotron 3.5 Transcribe), VAD, speaker diarization and noise suppression. It ships an OpenAI-compatible server and a live-transcription web page. No Python, no CUDA: Metal on Mac, Vulkan on Windows/Linux.

**Bottom line: it's an AI developer with a 10K following in Japanese dev circles, grown by shipping new models on day one and posting a video tweet with every release. The 11-01 wave came when Haas-ja-v1.0, a new Japanese speech recognition model, went viral and people wanted to run it on Windows; speech.cpp was nearly the only thing that could.** The differentiator is Japanese models + synthesis and recognition in one package + a single-file install with no Python + low first-audio latency. That's really why it took off, but it only holds within the small Japanese-speaking user base; globally it's modest for its category.

### 1. How the stars grew (data card github repo_day + GitHub releases + X)

| Timeline | Stars | Event |
|---|---|---|
| 09-29 | 0 | Created the repo, shipped v0.1.0 the same day |
| 10-09 | — | The author's long post 《speech.cpp を作ってみた話》 (Japanese/English, nyosegawa.com) |
| 10-14 | — | Release tweet for v0.8.5: "for AI assistants, fast TTS first-audio, light install, fast cold start, no Python/CUDA", 33.8K views, about 415 likes (x.com/gyakuse/status/2110192481161949264) |
| 10-14~10-21 | 129(10-16)→180(10-18) | Saved on Hatena Bookmark, 32 users, clustered in 10-15 (b.hatena.ne.jp/entry/s/github.com/nyosegawa/speech.cpp); 10-17 released v0.8.10 with the tweet "the CLI got cute"; 10-17 single-day +32 |
| 10-19~10-30 | 180→216 | Plateau, 0–5 stars per day |
| 10-21 | — | Two Zenn posts on 《Irodori-TTS の 7 実装を比較する》 that list speech.cpp as one of the implementations |
| 10-30~10-31 | 216→219 | Japanese company Third Intelligence released the Japanese ASR model Haas-ja-v1.0; speech.cpp shipped v0.9.3 with support the same day. A 10-31 tweet "adds Nemotron-3-Diarization, Nemotron-3.5-Transcribe-30B, Haas-ja-v1.0" with a video got 43.8K views, and replies came from English and Arabic speakers. |
| 10-31~11-03 | 219→324 | A batch of Qiita / ai-hardware-zukan / note / Zenn posts on "running Haas-ja on Windows", all using speech.cpp + GGUF. OpenHinoVoice (a Japanese local voice input tool, 51 stars) uses speech.cpp as its engine. 11-01 released v0.10.1 (NVIDIA RE-USE noise suppression), pinned tweet got 21K views; 11-01 single-day +60 (peak), 11-03 +31 |
| 11-02 | — | A Japanese Gigazine article on Haas-ja-v1.0 (didn't see speech.cpp mentioned; full text not verified) |
| 11-06 | 358 | Last 7 days +142, first 7 days +21 |

It never made GitHub trending (data card trending_days_any=0), and the Hacker News table has no posts about it (the hackernews table searches by title/repo, 0 results). No dedicated Reddit thread turned up. So the growth came almost entirely from **Japanese X and Japanese tech communities (Hatena, Zenn, Qiita, note)** . The stargazer timestamp endpoint wasn't available this time (GitHub returned 404), so the geographic distribution of stars is unverified.

### II. Competitors (doing the same thing)

| Name | Where | Launched | Scale | How it differs from its peers |
|---|---|---|---|---|
| whisper.cpp | Open source | 2022-09 | 55K stars | Whisper recognition only, no synthesis; Japanese accuracy trails the newer models |
| sherpa-onnx | Open source | 2022-09 | 16K stars | ONNX route, covers phones/embedded devices; many models, but new Japanese models (Irodori, Haas) weren't picked up right away |
| transcribe.cpp (handy-computer, promoted by Mozilla.ai) | Open source | 2026-04 | 2,139 stars; HN 773 points (2026-07-19) | Recognition only, 16+ model family; no TTS, no Japanese specialization |
| CrispASR | Open source | 2026-03 | 798 stars | A "speech model hub" forked from whisper.cpp with 24 ASR + 5 TTS models, broader than speech.cpp; Japanese models and first-audio latency aren't the focus |
| NVIDIA NeMo-Speech.cpp | Big-vendor open source | 2026-07 | 242 stars | NVIDIA's official ggml runtime, only runs the Nemotron family; speech.cpp also brings in Nemotron and adds Qwen and Japanese models |
| qwentts.cpp / predict-woo qwen3-tts.cpp / llama.cpp mainline Qwen3-TTS | Open source | 2026-02~05 | 200–250 stars; llama.cpp mainline already supports Qwen3-TTS voice cloning | A single-model TTS port; speech.cpp is synthesis + recognition + VAD in one |
| mlx-audio | Open source | 2024-11 | 8.2K stars | Apple MLX only; speech.cpp runs across Mac/Windows/Linux |
| Irodori-TTS official (Aratako) | Open source | 2026-02 | 1,547 stars | PyTorch reference implementation; Zenn measured a median M4 Pro first-audio of 1.497s, speech.cpp 0.214s |
| Speaches / Kokoro-FastAPI / LocalAI | Open source | 2023–2024 | 3.8K / 5.6K / 50K stars | Also offers an OpenAI-compatible speech API, but needs Python/Docker and uses different models |

No competitors at all is impossible; this space has been crowded for 2026 years. Its position is: **the only project that packages the new Japanese models (Irodori-TTS, ReazonSpeech, parakeet-ja, Haas-ja) together with Qwen3/Nemotron into one single-file, no-Python tool** .

### 3. Growth tactics (evidence)

1. **The author's own X account (confirmed)** : @gyakuse with 10K followers, a known individual developer in Japan's AI scene (previously notion-cli with 57 stars and agents-md-generator with 38 stars). Each major release gets a tweet with a screenshot or video: 10-14 with 33.8K views, 10-31 with 43.8K, 11-01 with 21K. This is the main engine.

2. **Hatena Bookmark (confirmed)** : 10-14–10-16 saves, 32 users, matching the first stretch of growth from 10-16 to 10-18.

3. **Riding the new model launch (confirmed)** : Haas-ja-v1.0 launched around 10-30, with only a PyTorch build officially. speech.cpp shipped GGUF and support the same day, and Japanese bloggers writing "how to run Haas-ja on Windows" used it (Qiita tasekino, ai-hardware-zukan, Zenn airtanker, note senfxn). So the new model's traffic flowed through tutorials to speech.cpp. 11-01 single-day +60 landed in the same window.

4. **Downstream project citing it (confirmed)** : OpenHinoVoice (iiinno, 51 stars) uses speech.cpp as its recognition engine, with a speech.cpp install one-liner as the first install step; the author's own ASIST (a desktop voice assistant) also uses it.

5. **Blog + self-published HF model repo (confirmed)** : a long bilingual Japanese/English post on nyosegawa.com; the author converts all models to GGUF and posts them on HF (sakasegawa/*), and the commands download them automatically by name. The HF model pages are an entry point too (traffic not verified).

6. **Third-party benchmark article (confirmed)** : someone on Zenn benchmarked it against the official implementation and reported "7x faster first-audio", which vouches for it.

7. No evidence found of HN posts, dedicated Reddit threads, GitHub trending, or English-language influencer shares.

### 4. Why it made the list

- The author's 10K-follower distribution in Japan's AI scene drives the traffic: each release tweet gets 2–40K views, and the first 129 stars came from the blog, tweets and Hatena together.

- It caught a gap: Japanese speech models only shipped as Python builds (Irodori-TTS, Haas-ja and others are PyTorch), which is painful for Windows users. speech.cpp offers a one-line install, pulls models by name, and picks up new models on day one.

- The second wave (starting 11-01) was clearly driven by Haas-ja-v1.0 going viral: Japanese bloggers wanted to try Haas-ja on Windows, and the tutorials all used speech.cpp.

- Release cadence is extremely fast: 38 releases in 29 days (v0.1.0→v0.10.1), and each version comes with new models to tweet about, keeping a steady supply of content.

- Positioning is clear: "built for voice assistants" — first-audio 0.029s (Qwen3-TTS 0.6B, M5), fast cold start, OpenAI-compatible API, just swap the base URL.

### 5. Differentiation

It doesn't have more models than whisper.cpp / sherpa-onnx / transcribe.cpp / CrispASR. What it does: (1) synthesis, recognition, VAD, diarization, and denoising in one binary; (2) first to support the Japanese models, and the author converts them to GGUF personally; (3) makes first-audio latency the selling point, with a web demo and live transcription; (4) checks each stage against the official implementation. **Japanese model support plus no-Python install is the direct reason it took off** ; the other points are reasons developers stay. Outside the Japanese community, the differentiation is thin (CrispASR and transcribe.cpp are better known), so stars have plateaued at 300+ within Japan.

### 6. Visibility and copycats

- Visibility: **it's circulating in small circles** — Japanese X, 32 Hatena Bookmark users, a handful of Zenn/Qiita tutorials. No HN, no GitHub trending, and I haven't seen English-language accounts or media covering it (Gigazine covered the Haas-ja model itself).

- Copycats: I haven't seen any project directly copying it. Some downstream projects are built on it (OpenHinoVoice). "xxx.cpp speech engine" projects were already common, so it isn't really being imitated.

### VII. What others can borrow

- How local AI tools win: **watch for new open-source models in one language or region, and ship a "run it with one command" version the day they drop** . Tutorial writers will distribute it for you.

- Pair each release with a new model and a tweet with a video, so the release itself is the content.

- Convert the model yourself, host it on HF, and let commands download by name, eliminating the environment-setup step.

- Anchor the pitch on a measurable number (first audio 0.029s / 7x faster) so others can write comparison posts.

### 8. What's still uncertain

- Stargazer geography and timestamps weren't retrieved (GitHub API 404). "Mostly Japanese users" is inferred from the channels.

- Of 11-01 stars on 60, we can't tell how many came from 10-31 tweets (with English replies) and how many from Haas-ja's tutorial.

- Unverified whether NVIDIA-related accounts retweeted the Nemotron support post.

- Actual usage (release download counts, HF model downloads) was not checked.

### Sources

- Repo and releases: [github.com/nyosegawa/speech.cpp](https://github.com/nyosegawa/speech.cpp); GitHub API (2026-11-06)

- Data card daily stars: our github repo_day

- Tweets: x.com/gyakuse/status/2110192481161949264 (10-14,33.8K views), /2116322257454911772 (10-31,43.8K), /2116928286286901601 (11-01,21K); profile x.com/gyakuse (10K followers)

- はてブ:[b.hatena.ne.jp/entry/s/github.com/nyosegawa/speech.cpp](https://b.hatena.ne.jp/entry/s/github.com/nyosegawa/speech.cpp)(32 users)

- Blog: [nyosegawa.com/posts/speech-cpp/](https://nyosegawa.com/posts/speech-cpp/), [nyosegawa.com/en/posts/speech-cpp-weight-types-and-ggml-fork…](https://nyosegawa.com/en/posts/speech-cpp-weight-types-and-ggml-fork/)

- Haas-ja related: [qiita.com/tasekino/items/99141a0e287126ec5ffc](https://qiita.com/tasekino/items/99141a0e287126ec5ffc), [ai-hardware-zukan.com/haas-ja-asr-speechcpp-whisper-benchmar…](https://ai-hardware-zukan.com/haas-ja-asr-speechcpp-whisper-benchmark/), [zenn.dev/airtanker/articles/haas-ja-mac-dictation](https://zenn.dev/airtanker/articles/haas-ja-mac-dictation), [note.com/senfxn/n/nf03aaece8962](https://note.com/senfxn/n/nf03aaece8962), [gigazine.net/news/20261102-haas-ja-v1-asr/](https://gigazine.net/news/20261102-haas-ja-v1-asr/)

- OpenHinoVoice: [aegisfleet.github.io/tool-survey-report/reports/open-hino-vo…](https://aegisfleet.github.io/tool-survey-report/reports/open-hino-voice/), github.com/iiinno/OpenHinoVoice

- Zenn reviews: [zenn.dev/yoshitetsu/articles/speech-cpp-irodori-m4-pro](https://zenn.dev/yoshitetsu/articles/speech-cpp-irodori-m4-pro), [zenn.dev/idunaf/articles/215bfb63b29db0](https://zenn.dev/idunaf/articles/215bfb63b29db0)

- Competitor stars: GitHub API 2026-11-06; transcribe.cpp on HN: hackernews library (2026-07-19,773 points)

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