Demo

The full demo recording shows SoCa's reviewed private-vault workflow from spoken request to grounded response.

Full demo recording

Short project teaser

Abstract

SoCa is a Vietnamese voice agent designed around a local audio path and an explicitly selected local or remote language model. Its runtime separates microphone input, VAD/AEC, ASR, a controlled workflow, tools and evidence-aware knowledge or memory, the selected LLM, and TTS playback. The knowledge path combines sparse retrieval with dense representations and keeps evidence, citations, and terminal outcomes observable instead of turning a weak match into a confident answer. Evaluation uses public speech and retrieval datasets together with private release manifests, real echo recordings, and reviewed conversational trajectories. The page reports what was measured, what remains blocked, and which boundaries are deliberate; it does not present a paper or claim general voice-assistant capability.

Benchmark evidence

Three measured views of SoCa's core runtime.

Qwen3-ASR release qualification matrix comparing WER, latency, and resource measurements — **ASR release qualification.** Qwen3-ASR 0.6B and 1.7B; pinned FLEURS-vi, private code-switch, and non-speech manifests; production qualification run in 2026.

Results

Values below are copied from the SoCa benchmark record. Dataset revisions, commands, hardware, and unresolved gates remain in BENCHMARKS.md.

Selected SoCa benchmark results
Metric	Value	n	Dataset / measured	Date
Qwen release FLEURS-vi WER	8.59%	30 utterances	FLEURS vi_vn production paired run	2026-08-02
Qwen release cold inference p95	691 ms	release matrix	qwen-release on Apple Silicon	2026-08-02
Retrieval Recall@5	0.9161	1,000 queries	TVPL · fusion selection	2026-07-30
Retrieval p95 per query	71.019 ms	1,000 queries	TVPL · production fusion candidate	2026-07-30
Unanswerable false evidence	0 / 8	20 questions	XQuAD grounding split	2026-07-31
Barge-in false interrupt	2.7%	300 pairs	AEC-Challenge real echo	2026-07-29
Barge-in detection	94.7%	300 pairs	AEC-Challenge real echo	2026-07-29
Turn-taking premature close	18.3%	120 timelines	Vietnamese turn-taking benchmark	2026-07-29

Limitations

Goal-level routing can still choose the wrong capability tier, so retrieval is not treated as proof that an end-to-end answer is correct; the real-vault false-evidence gate remains visible in the benchmark record.
Local audio qualification is currently centered on Apple Silicon; ARM/Linux and a complete hardware-microphone release run are not claimed here.
ASR confidence calibration is profile-specific. A model without its own calibration is not silently substituted with another model.
The production tool surface is intentionally small. Device control, weather, and broad external actions are outside this release page.

BibTeX

@software{vo2026soca,
  author  = {Võ Minh Thịnh},
  title   = {SoCa: A Vietnamese Voice Agent with Evidence-Aware Retrieval},
  year    = {2026},
  url     = {https://github.com/finalflash159/soca}
}