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
Evaluation results
100%
68%Telco Contact-Center Live Transcription Blueprint
Romanian contact-center real-time ASR blueprint
Blueprint sections present
25%
100%
Romanian streaming model
0%
100%
SimulStreaming or AlignAtt named
0%
100%
Silero VAD included
100%
100%
Romanian alignment heads note
0%
100%
Chunk size specified
16%
100%
Telephony normalization
100%
100%
70% headroom sizing math
12%
100%
Per-stream VRAM stated
16%
100%
Romanian eval benchmarks
28%
100%
Hallucination probe
0%
100%
Rejected Romanian-excluded models
0%
100%
Model pick has license
100%
100%
No implementation artifacts
0%
100%
No CPU/memory autoscaling
100%
100%
100%
37%Customer Experience Platform: Large-Scale English Transcription Blueprint
High-concurrency Whisper serving architecture
Triton + TensorRT-LLM recommended
30%
100%
Decoupled streaming mode
100%
100%
Bidirectional gRPC required
71%
100%
Inflight fused batching
0%
100%
KV cache tuning note
50%
100%
Replica scaling with stream affinity
100%
100%
Consistent hashing for affinity
100%
100%
70% headroom GPU sizing
62%
100%
H100 per-GPU concurrency anchor
33%
100%
TensorRT-LLM runtime with FP16 or FP8
0%
100%
Queue-based autoscaling metric
100%
100%
No FastAPI as inference engine
100%
100%
No implementation artifacts
100%
100%
Evaluation plan present
20%
100%
100%
45%Bilingual Support Transcription with English Translation
RO+EN code-switching with English translation
No forced language=ro
80%
100%
Canary-1B-v2 for translation
0%
100%
Whisper-turbo translation rejected
0%
100%
MMS or SeamlessM4T-v2 excluded
77%
100%
Per-segment LID for code-switching
87%
100%
Canary license stated
0%
100%
Moldovan dialect as research gap
50%
100%
Romanian eval benchmarks
0%
100%
CER for Romanian
100%
100%
Blueprint sections present
71%
100%
Rejected alternatives with reasons
100%
100%
Runtime appropriate for L4
83%
100%
No implementation artifacts
100%
100%
- Evaluated
- Agent
- Claude Code
- Model
- Claude Sonnet 4.6