HN Showcase
Back to browse
17MB pronunciation scorer beats human experts at phoneme level

17MB pronunciation scorer beats human experts at phoneme level

by fabiosuizu·Feb 20, 2026·4 points·2 comments
Visit ProjectView on HN

AI Analysis

●●●BangerWizardryBig BrainSolve My Problem

Beats human experts at phoneme scoring while 70x smaller than SOTA models.

Strengths
  • •Quantized Citrinet-256 backbone achieves superhuman inter-annotator agreement (0.580 PCC vs 0.555 human) at 17MB—genuine constraint engineering.
  • •Sub-300ms CPU latency + REST/MCP/Azure APIs make it production-viable for real-time language learning feedback loops.
  • •CTC forced alignment + GOP scoring + ensemble methodology is technically non-obvious and well-benchmarked on standard datasets.
Weaknesses
  • •10-15% raw accuracy gap vs SOTA suggests limited upside for high-precision use cases beyond language learning apps.
  • •Demo UI is functional but barebones—no clear onboarding for developers to understand when/why to use this vs larger models.
Category
AI/ML
Target Audience

Language learning app developers, speech-enabled education platforms, ESL instructors

Similar To

Google Cloud Speech-to-Text (pronunciation confidence scoring) · Speechocean762 reference systems (wav2vec2-based) · Azure Speech Service (pronunciation assessment)

Post Description

I built an English pronunciation assessment engine that fits in 17MB and runs in under 300ms on CPU.

Architecture: CTC forced alignment + GOP scoring + ensemble heads (MLP + XGBoost). No wav2vec2 or large self-supervised models — the entire pipeline uses a quantized NeMo Citrinet-256 as the acoustic backbone.

Benchmarked on speechocean762 (standard academic benchmark, 2500 utterances): - Phone accuracy (PCC): 0.580 — exceeds human inter-annotator agreement (0.555) - Sentence accuracy: 0.710 — exceeds human agreement (0.675) - Model is 70x smaller than wav2vec2-based SOTA

Trade-off: we're ~10-15% below SOTA on raw accuracy. But for real-time feedback in language learning apps, the latency/size trade-off is worth it.

Available as REST API, MCP server (for AI agents), and on Azure Marketplace.

Demo: https://huggingface.co/spaces/fabiosuizu/pronunciation-asses...

Interested in feedback on the scoring approach and use cases people would find valuable.

Similar Projects

AI/ML●Mid

Darius – An AI router that selects the best model for each prompt

The product puts model selection behind a friendly chat UI — I can see model tags like XAI:GROK-4-1-FAST-REASONING in the screenshots — and leans hard on privacy and 'no dark patterns' messaging. The UX is clean and approachable, but the routing logic is opaque and this sits in a crowded space of multi-LLM frontends (Poe, Perplexity, etc.), so the value depends on how smart and cost-effective their orchestration actually is.

SlickSolve My Problem
mazenkurdi
304mo ago