niedev/RTranslator

Worth trying if you own a reasonably recent Android phone and want conversation translation without sending audio to a cloud service; otherwise the hardware and licensing caveats are the deciding factors.

The problem

Before, a face-to-face conversation across languages meant typing into a cloud translator and passing a phone back and forth, or trusting a shared connection. That breaks down for longer exchanges, for multiple participants, and where connectivity is poor. RTranslator moves recognition and translation onto the phones themselves: one phone captures speech, transcribes it, sends the text over a local connection to the other phone, which translates it and reads it aloud. The result is a back-and-forth conversation where each person hears the other in their own language, whether the screen is on or the phone is in a pocket.

What's different

The distinguishing piece is the Conversation mode: two or more phones pair, each side uses its own language model locally, and only text crosses the link. Bluetooth headsets can be used instead of the phone microphone and speaker, which is what makes a pocketed-phone conversation plausible. WalkieTalkie mode covers the short encounter — asking directions, talking to a shop assistant — with a single phone that listens for two selectable languages, detects which is being spoken, and speaks the translation aloud before resuming listening. There is also a plain text translator.

The trade-off is that local inference costs memory and time. The README reports its quantized NLLB build at roughly 1.3 GB RAM and about 2 seconds for 75 tokens, and its Whisper build at 0.9 GB and 1.6 seconds for 11 seconds of audio. The 3.0 beta switches to selectable Mozilla Bergamot, Madlad 400 3B, and HY-MT 1.5 1.8B models, and replaces MLKit with FastText and Lingua to make the app fully open source — but that beta is still beta. The "almost open-source" hedge is about model licensing, not the app code.

Who should try it

Good fit: a traveler or aid worker with a 6 GB-plus Android phone who needs repeated two-way conversation in areas with weak or no data, and who is comfortable installing an APK from GitHub releases and letting it download about 1.2 GB of models on first launch. Also a reasonable fit if you prefer not to route your speech through a hosted service.

Poor fit: iPhone users, anyone expecting a turnkey Play Store install today, and people on low-RAM devices who need fast responses. It is also a poor fit for someone who needs guaranteed translation quality across many language pairs — the model choices in the 3.0 beta differ in coverage and quality, and the README's quality claims are the author's own.

Requirements and limitations

The practical floor is RAM: the README says at least 6 GB to avoid crashes, and the author states in a 2024 HN reply that slower CPUs mainly increase latency rather than making the app unusable (author's reply). Conversation mode needs a second phone running the app; WalkieTalkie is the single-device path. Bluetooth connections are documented as occasionally dropping, and the README warns that background power limits on some phones make keeping the app open the safer habit.

Setup has a real dependency on GitHub: the first-run model download comes from there, and the README points people in slow-GitHub regions to a manual model sideloading guide instead (sideloading instructions). Licensing is the other boundary. NLLB was CC-BY-NC, and an HN commenter in June 2024 argued that makes the project not open source under OSI's definition (licensing thread); the 3.0 beta's model swap is the project's answer. If license terms matter to your use, check the current model licenses rather than the app's Apache-2.0 code license.

HotRepo's take

HotRepo recorded one daily-board appearance on 2026-10-26 at 10,552 stars, +13 that day; there is no comparable series, so persistence of interest cannot be judged from it. The 2024 Show HN thread drew 236 points and 56 comments (discussion), but that is discussion volume, not a quality signal. Before committing, verify that the 3.0 beta's chosen model supports your language pair and fits your phone's RAM.