sharaf/speech-recognition-architect
Use when the user wants to design, size, audit, or choose a self-hosted speech recognition or streaming ASR stack, including Whisper, Parakeet, Canary, Riva, NIM, Triton ASR, faster-whisper, sherpa-onnx, voice-agent transcription, Romanian or Moldovan ASR, contact-center transcription, GPU sizing, latency budgets, multilingual routing, VAD, diarization, or production evaluation.
100
100%
Does it follow best practices?
Impact
100%
2.00xAverage score across 3 eval scenarios
Passed
No known issues
task.mdevals/scenario-3/
Bilingual Support Transcription with English Translation
Problem/Feature Description
A European legal-tech SaaS company serves clients across Romania and the Republic of Moldova. Their AI-assisted contract review platform captures audio from client intake sessions where lawyers and clients speak freely in a mix of Romanian and English — sometimes switching mid-sentence. Additionally, some callers from the Republic of Moldova speak a regional dialect. The company's back-end analytics pipeline expects all transcripts to arrive as English text, so every non-English segment must be translated to English, not just transliterated.
The product manager has specified that the platform must be commercially deployable without per-user or per-transcript royalties. The development team is exploring models they have heard about — including some research models with excellent multilingual coverage — but they have not committed to any specific stack. They operate modest on-premises infrastructure: a single NVIDIA L4 GPU node with 24 GB VRAM, targeting a batch-style processing mode where audio is submitted as completed session recordings (not real-time streams), and an average session is 15-20 minutes. There are rarely more than 30 sessions being processed simultaneously.
Produce a complete architecture blueprint for this system. Address how the system should handle the bilingual audio, what model or models best serve translation to English, what runtime and precision are appropriate for the hardware, how processing layers should be composed, what the evaluation plan should include for this language pair and dialect situation, and what risks the team should be aware of.
Output Specification
Write the blueprint as a single Markdown file named blueprint.md. The document should make deliberate technology choices and explain what alternatives were considered and why they were not selected. The evaluation plan must address both the Romanian and English portions of the audio and account for the Moldovan dialect situation. Include a hardware sizing calculation.