Voice Reach Running Notes

Community Article
Published June 14, 2026

Voice Reach V1 is the Build Small Hackathon version of the metime.to voice contact widget: an embeddable contact flow where visitors speak in Hindi or Hinglish, review an editable English message, and send it through a site-owner demo flow.

Artifacts

Current State

The current public Article should be updated from this file and treated as the single running-notes Article for Voice Reach. The older supporting Article URL remains published provenance, but its content now belongs here instead of in a separate local article source.

Hosted proof should be read narrowly: Voice Reach has a private Space package, trace surfaces, and recorded model/runtime smoke evidence. A fresh public submission claim still needs a current hosted run with fallback_used=false.

Notes

2026-06-15 - Model And Runtime Note

The current decision signal is Nemotron ASR plus explicit language hint for ASR, with tiny-aya fire as the best current Hindi/Hinglish text default. MiniCPM5 remains useful for English smoke/default fallback. Cohere official ASR should stay deferred until a comparison smoke is worth a paid Modal redeploy and live run.

Pinned artifacts from the original supporting-note publication:

Model decision snapshot:

Role Model Runtime Shape Finding
ASR onnx-community/nemotron-3.5-asr-streaming-0.6b-onnx-int4 ONNX / NeMo-style ASR path Best current ASR signal with explicit language hint.
ASR onnx-community/cohere-transcribe-03-2026-ONNX Intended Transformers.js/WebGPU; current Python path unsupported Selector-visible, but current Modal endpoint returns an explicit adapter blocker instead of fake output.
ASR CohereLabs/cohere-transcribe-03-2026 Transformers Adapter call shape was corrected, but no redeployed live proof is counted yet.
Text openbmb/MiniCPM5-1B-GGUF llama.cpp / GGUF Runnable hosted proof exists; useful as English smoke/default fallback.
Text CohereLabs/tiny-aya-fire-GGUF llama.cpp / GGUF Best current Hindi/Hinglish text signal, with higher latency.
Text nvidia/NVIDIA-Nemotron-3-Nano-4B-GGUF llama.cpp / GGUF Runnable but weak on the current seed set.

What ran:

  • Corrected the official Cohere ASR comparison adapter to the HF model-card call shape.
  • Generated a privacy-safe agent trace bundle with 36 hosted trace rows.
  • Uploaded the private Space to build-small-hackathon/voice-reach.
  • Uploaded the private trace dataset to build-small-hackathon/voice-reach-agent-traces.

What this proves:

  • The Article/dataset bundle carries privacy-safe hosted trace rows and schema.
  • Nemotron ASR with explicit language hint is the best current ASR signal.
  • tiny-aya fire is the best current Hindi/Hinglish text signal.

What this does not prove:

  • Broad ASR quality across real traffic.
  • Production email delivery.
  • Judge-ready status.
  • Cohere official ASR proof after the adapter correction.

Source Provenance

These paths are in the source repo, not necessarily in the Space package:

  • product/5-idea-voice-contact-widget/v1/space/
  • product/5-idea-voice-contact-widget/v1/evidence/runtime-target-matrix-modal-smoke-2026-06-15.md
  • product/5-idea-voice-contact-widget/v1/evidence/voice-reach-v1-hf-hackathon-boot-2026-06-15.md
  • product/5-idea-voice-contact-widget/v0/traces_public/data/eval_traces.jsonl
  • product/5-idea-voice-contact-widget/v0/eval/results/quick-signals.json

Community

Sign up or log in to comment