PanorVoice-Refine-V1

Model purpose

A local ASR transcript refinement model developed for PanorVoice. It cleans raw speech-to-text output (punctuation, capitalization, spoken commands, filler cleanup) while preserving meaning and intent.

Supported task

• Input: raw ASR-like transcript text
• Output: cleaned transcript text
• Not for: open-ended chat, factual QA, or inventing content that was not spoken

Base model

Fine-tuned from: Qwen/Qwen3.5-2B-Base

Exact tokenizer/config files ship with the packaged weights and must be attributed per the base-model license (see LICENSE and THIRD_PARTY_NOTICES.md).

Training methodology

• Method: LoRA fine-tuning via MLX-LM on Apple Silicon
• Objective: map degraded/spoken ASR text → clean transcript
• Adapters may be fused into a standalone export for local inference

Dataset methodology

Synthetic ASR degradation of clean demonstration/safe text (spoken punctuation, fillers, repetitions, self-corrections, technical vocabulary coverage). No proprietary third-party speech corpora are redistributed here.

Inference

Intended for on-device use with PanorVoice (MLX on macOS Apple silicon). Apply deterministic spoken-command preprocessing before model inference when possible.

Example:

Input:  hey john comma the playwright tests are completed period
Output: Hey John, the Playwright tests are completed.

Limitations

• May imperfectly restore rare proper nouns without vocabulary hints
• Must not be trusted to add facts that were not in the ASR input
• Negation and intent preservation are hard requirements; verify on safety suites
• Windows/Linux inference backends may differ from the macOS MLX path

Known limitations

• Demo/synthetic training data is not a substitute for production-scale corpora
• Quantized (4-bit) exports may slightly reduce formatting fidelity

License

License status for this package: Apache-2.0

Apache License 2.0 (see LICENSE). Base-model attribution is in THIRD_PARTY_NOTICES.md.

Third-party notices

See THIRD_PARTY_NOTICES.md for base model, tokenizer, and tooling attributions.

Ethical / use limitations

• Do not use to fabricate transcripts of speech that did not occur
• Do not use to alter meaning in legal, medical, or compliance-critical contexts without human review
• Preserve user privacy: prefer local inference

Model version

0.1.0