1. Summary
ARconvo is an offline Arabic–English bilateral speech translator that runs on a single PC, built for bilingual workshops and transcript records. It detects which language is being spoken, transcribes it with Faster-Whisper, translates it with a local model, speaks the translation and records a bilingual transcript. Nothing leaves the PC, so it works without internet and keeps sensitive conversations private.
2. Problem and context
Who has the problem. Facilitators and participants in workshops and training sessions where some people speak Arabic and others speak English.
What it costs them.
- Assumption: without an interpreter, sessions slow down, people repeat themselves, and Arabic- or English-only participants miss content.
- Assumption: interpreters are not always available or affordable for routine sessions.
- Cloud translation services need a connection and send the conversation to a third party, which rules them out in low-connectivity or privacy-sensitive settings.
- Workshops often need a record of what was said in both languages.
3. Users and personas
| Persona | Need |
|---|---|
| Facilitator or trainer | Speak naturally in their own language and be understood |
| Participant | Hear and read what was said in their own language |
| Record keeper | A bilingual transcript of the session |
| Native Arabic-speaking tester | Check detection, translation accuracy and interaction quality during evaluation |
4. Goals and non-goals
Goals
- Work fully offline on one PC.
- Translate both ways, Arabic to English and English to Arabic, with automatic language detection.
- Produce a bilingual transcript.
- Run on a PC with 4 GB of VRAM.
Non-goals
- A cloud service.
- Languages other than Arabic and English.
- Meeting-platform integration in the current version (planned, see section 9).
5. Requirements
Functional requirements
A-1. Speak in either language. As a speaker, I want to talk in Arabic or English without pressing a language button.
- The language is detected automatically for each utterance.
- Silence detection marks where an utterance ends.
A-2. Hear and read the translation. As a listener, I want to hear and read what was said in my language.
- Speech is transcribed, translated and spoken in the other language.
- English speech uses Windows-native text-to-speech; Arabic speech uses model-based synthesis.
A-3. Bilingual transcript. As a record keeper, I want the session captured in both languages.
- Arabic and English transcripts are shown side by side in live panels and captured for the record.
A-4. Control and status. As a facilitator, I want to control the session and see that it is working.
- Controls run in the browser on the same PC.
- Live system status shows what each part of the pipeline is doing.
Non-functional requirements
| Area | Requirement |
|---|---|
| Offline | Every step runs locally; no internet connection needed |
| Privacy | Audio, transcripts and translations stay on the PC |
| Hardware | Runs on a PC with 4 GB of VRAM |
| Performance | No published latency target yet; see "Metrics to track" |
| Language quality | Arabic dialect handling, validated with native Arabic speakers |
6. Solution overview and architecture
7. AI design
Models
| Stage | Model |
|---|---|
| Speech recognition | Faster-Whisper with an Arabic-dialect-trained model |
| Translation | Falcon 3B, run locally |
| Speech synthesis | Windows-native English TTS; model-based Arabic synthesis |
Model choice was constrained by the hardware: everything has to fit and run on a PC with 4 GB of VRAM.
Evaluation. Iterative evaluation with native Arabic-speaking testers, validated against three things: Arabic detection, translation accuracy and interaction quality.
Human in the loop. The people in the conversation are the check: they hear the translation and see both transcripts live, so they can catch and correct a mistranslation on the spot.
Failure handling. Live system status shows the facilitator when a stage is struggling. Assumption: there is no automatic fallback yet when detection or translation confidence is low; the speaker repeats or rephrases.
8. Success metrics
Confirmed
- None published yet. ARconvo is a working PC app.
Metrics to track
- End-to-end delay from the end of an utterance to the translated speech
- Language detection accuracy
- Word error rate for Arabic and English recognition
- Translation accuracy, rated by native speakers
- GPU memory headroom on a 4 GB card
- Sessions run and transcripts produced
9. Release history and roadmap
| When | Release |
|---|---|
| Current | Working offline PC app |
| Planned | Meeting-platform integration for bilingual workshops and training sessions |
10. Risks and open questions
Risks
- Dialect variation. Arabic dialects vary widely; accuracy may drop for dialects the model saw less of.
- Hardware limits. A 4 GB VRAM budget caps model size, and with it, translation quality.
- Overlapping speech. Assumption: the pipeline expects one speaker at a time.
- Over-reliance. Machine translation should not be the only record in high-stakes settings without human review.
Open questions
- How should low-confidence detection or translation be shown to listeners?
- Which meeting platform should the integration target first?
- Should transcripts be exportable in a standard subtitle or document format?
11. My role
- Builder of ARconvo.
- Assumption: that covers the offline-first, privacy-first design, choosing models that fit a 4 GB VRAM PC, and running the evaluation with native Arabic-speaking testers.