Improve voice recognition and correction quality

[IchenDEV]

Summary

• improve Apple Speech, WhisperKit, and Qwen recognition paths with contextual vocabulary, safer model selection, decoding fallbacks, VAD, and high-quality 16 kHz mono preprocessing
• make LLM correction evidence-bounded through request snapshots, faithful prompts, truncation handling, protected-token checks, semantic fidelity checks, and safe fallback to prepared ASR text
• improve dictionary replacement behavior, deferred replacement safety, and OCR-backed screen context
• add regression coverage plus a dependency-free CER/WER, terminology, entity-fidelity, hallucination, and latency evaluator
• document the research findings, open-source references, rollout plan, and remaining accuracy gaps

Why

Recognition quality was being reduced by inconsistent language and dictionary context, brittle Whisper model resolution, unsuitable Qwen input audio, and overly shallow decoding fallbacks. The correction layer could also accept truncated or semantically altered LLM output, while the project lacked a repeatable way to measure raw ASR quality separately from post-processing quality.

Impact

Users should get more robust recognition across Apple Speech, Whisper, and Qwen paths, with better handling of personal vocabulary and fewer correction-stage changes to numbers, dates, units, URLs, email addresses, file paths, quoted text, polarity, and self-corrections. When correction cannot be proven faithful, OpenType now preserves the prepared recognition result instead of inserting a risky rewrite.

The included example corpus validates the evaluator pipeline only. It is not evidence of production accuracy improvement; real-device recordings are still required for comparative benchmarking.

Validation

• swift test: 529 XCTest tests passed, 7 environment-dependent tests skipped
• Swift Testing: 2 tests passed
• Python evaluator tests: 6 passed
• bash scripts/ci-basic-checks.sh: passed
• staged diff and conflict-marker checks: passed

Research

See docs/superpowers/specs/2026-07-30-voice-quality-research.md for source evaluation, licensing notes, phased recommendations, and the proposed real-corpus benchmark.

[IchenDEV]

Summary

• improve Apple Speech, WhisperKit, and Qwen recognition paths with contextual vocabulary, safer model selection, decoding fallbacks, VAD, and high-quality 16 kHz mono preprocessing
• make LLM correction evidence-bounded through request snapshots, faithful prompts, truncation handling, protected-token checks, semantic fidelity checks, and safe fallback to prepared ASR text
• improve dictionary replacement behavior, deferred replacement safety, and OCR-backed screen context
• add regression coverage plus a dependency-free CER/WER, terminology, entity-fidelity, hallucination, and latency evaluator
• document the research findings, open-source references, rollout plan, and remaining accuracy gaps

Why

Recognition quality was being reduced by inconsistent language and dictionary context, brittle Whisper model resolution, unsuitable Qwen input audio, and overly shallow decoding fallbacks. The correction layer could also accept truncated or semantically altered LLM output, while the project lacked a repeatable way to measure raw ASR quality separately from post-processing quality.

Impact

Users should get more robust recognition across Apple Speech, Whisper, and Qwen paths, with better handling of personal vocabulary and fewer correction-stage changes to numbers, dates, units, URLs, email addresses, file paths, quoted text, polarity, and self-corrections. When correction cannot be proven faithful, OpenType now preserves the prepared recognition result instead of inserting a risky rewrite.

The included example corpus validates the evaluator pipeline only. It is not evidence of production accuracy improvement; real-device recordings are still required for comparative benchmarking.

Validation

• swift test: 529 XCTest tests passed, 7 environment-dependent tests skipped
• Swift Testing: 2 tests passed
• Python evaluator tests: 6 passed
• bash scripts/ci-basic-checks.sh: passed
• staged diff and conflict-marker checks: passed

Research

See docs/superpowers/specs/2026-07-30-voice-quality-research.md for source evaluation, licensing notes, phased recommendations, and the proposed real-corpus benchmark.

[IchenDEV]

Summary

• improve Apple Speech, WhisperKit, and Qwen recognition paths with contextual vocabulary, safer model selection, decoding fallbacks, VAD, and high-quality 16 kHz mono preprocessing
• make LLM correction evidence-bounded through request snapshots, faithful prompts, truncation handling, protected-token checks, semantic fidelity checks, and safe fallback to prepared ASR text
• improve dictionary replacement behavior, deferred replacement safety, and OCR-backed screen context
• add regression coverage plus a dependency-free CER/WER, terminology, entity-fidelity, hallucination, and latency evaluator
• document the research findings, open-source references, rollout plan, and remaining accuracy gaps

Why

Recognition quality was being reduced by inconsistent language and dictionary context, brittle Whisper model resolution, unsuitable Qwen input audio, and overly shallow decoding fallbacks. The correction layer could also accept truncated or semantically altered LLM output, while the project lacked a repeatable way to measure raw ASR quality separately from post-processing quality.

Impact

Users should get more robust recognition across Apple Speech, Whisper, and Qwen paths, with better handling of personal vocabulary and fewer correction-stage changes to numbers, dates, units, URLs, email addresses, file paths, quoted text, polarity, and self-corrections. When correction cannot be proven faithful, OpenType now preserves the prepared recognition result instead of inserting a risky rewrite.

The included example corpus validates the evaluator pipeline only. It is not evidence of production accuracy improvement; real-device recordings are still required for comparative benchmarking.

Validation

• swift test: 529 XCTest tests passed, 7 environment-dependent tests skipped
• Swift Testing: 2 tests passed
• Python evaluator tests: 6 passed
• bash scripts/ci-basic-checks.sh: passed
• staged diff and conflict-marker checks: passed

Research

See docs/superpowers/specs/2026-07-30-voice-quality-research.md for source evaluation, licensing notes, phased recommendations, and the proposed real-corpus benchmark.

[IchenDEV]

Summary

• improve Apple Speech, WhisperKit, and Qwen recognition paths with contextual vocabulary, safer model selection, decoding fallbacks, VAD, and high-quality 16 kHz mono preprocessing
• make LLM correction evidence-bounded through request snapshots, faithful prompts, truncation handling, protected-token checks, semantic fidelity checks, and safe fallback to prepared ASR text
• improve dictionary replacement behavior, deferred replacement safety, and OCR-backed screen context
• add regression coverage plus a dependency-free CER/WER, terminology, entity-fidelity, hallucination, and latency evaluator
• document the research findings, open-source references, rollout plan, and remaining accuracy gaps

Why

Recognition quality was being reduced by inconsistent language and dictionary context, brittle Whisper model resolution, unsuitable Qwen input audio, and overly shallow decoding fallbacks. The correction layer could also accept truncated or semantically altered LLM output, while the project lacked a repeatable way to measure raw ASR quality separately from post-processing quality.

Impact

Users should get more robust recognition across Apple Speech, Whisper, and Qwen paths, with better handling of personal vocabulary and fewer correction-stage changes to numbers, dates, units, URLs, email addresses, file paths, quoted text, polarity, and self-corrections. When correction cannot be proven faithful, OpenType now preserves the prepared recognition result instead of inserting a risky rewrite.

The included example corpus validates the evaluator pipeline only. It is not evidence of production accuracy improvement; real-device recordings are still required for comparative benchmarking.

Validation

• swift test: 529 XCTest tests passed, 7 environment-dependent tests skipped
• Swift Testing: 2 tests passed
• Python evaluator tests: 6 passed
• bash scripts/ci-basic-checks.sh: passed
• staged diff and conflict-marker checks: passed

Research

See docs/superpowers/specs/2026-07-30-voice-quality-research.md for source evaluation, licensing notes, phased recommendations, and the proposed real-corpus benchmark.

[IchenDEV]

Summary

• improve Apple Speech, WhisperKit, and Qwen recognition paths with contextual vocabulary, safer model selection, decoding fallbacks, VAD, and high-quality 16 kHz mono preprocessing
• make LLM correction evidence-bounded through request snapshots, faithful prompts, truncation handling, protected-token checks, semantic fidelity checks, and safe fallback to prepared ASR text
• improve dictionary replacement behavior, deferred replacement safety, and OCR-backed screen context
• add regression coverage plus a dependency-free CER/WER, terminology, entity-fidelity, hallucination, and latency evaluator
• document the research findings, open-source references, rollout plan, and remaining accuracy gaps

Why

Recognition quality was being reduced by inconsistent language and dictionary context, brittle Whisper model resolution, unsuitable Qwen input audio, and overly shallow decoding fallbacks. The correction layer could also accept truncated or semantically altered LLM output, while the project lacked a repeatable way to measure raw ASR quality separately from post-processing quality.

Impact

Users should get more robust recognition across Apple Speech, Whisper, and Qwen paths, with better handling of personal vocabulary and fewer correction-stage changes to numbers, dates, units, URLs, email addresses, file paths, quoted text, polarity, and self-corrections. When correction cannot be proven faithful, OpenType now preserves the prepared recognition result instead of inserting a risky rewrite.

The included example corpus validates the evaluator pipeline only. It is not evidence of production accuracy improvement; real-device recordings are still required for comparative benchmarking.

Validation

• swift test: 529 XCTest tests passed, 7 environment-dependent tests skipped
• Swift Testing: 2 tests passed
• Python evaluator tests: 6 passed
• bash scripts/ci-basic-checks.sh: passed
• staged diff and conflict-marker checks: passed

Research

See docs/superpowers/specs/2026-07-30-voice-quality-research.md for source evaluation, licensing notes, phased recommendations, and the proposed real-corpus benchmark.

[IchenDEV]

Summary

• improve Apple Speech, WhisperKit, and Qwen recognition paths with contextual vocabulary, safer model selection, decoding fallbacks, VAD, and high-quality 16 kHz mono preprocessing
• make LLM correction evidence-bounded through request snapshots, faithful prompts, truncation handling, protected-token checks, semantic fidelity checks, and safe fallback to prepared ASR text
• improve dictionary replacement behavior, deferred replacement safety, and OCR-backed screen context
• add regression coverage plus a dependency-free CER/WER, terminology, entity-fidelity, hallucination, and latency evaluator
• document the research findings, open-source references, rollout plan, and remaining accuracy gaps

Why

Recognition quality was being reduced by inconsistent language and dictionary context, brittle Whisper model resolution, unsuitable Qwen input audio, and overly shallow decoding fallbacks. The correction layer could also accept truncated or semantically altered LLM output, while the project lacked a repeatable way to measure raw ASR quality separately from post-processing quality.

Impact

Users should get more robust recognition across Apple Speech, Whisper, and Qwen paths, with better handling of personal vocabulary and fewer correction-stage changes to numbers, dates, units, URLs, email addresses, file paths, quoted text, polarity, and self-corrections. When correction cannot be proven faithful, OpenType now preserves the prepared recognition result instead of inserting a risky rewrite.

The included example corpus validates the evaluator pipeline only. It is not evidence of production accuracy improvement; real-device recordings are still required for comparative benchmarking.

Validation

• swift test: 529 XCTest tests passed, 7 environment-dependent tests skipped
• Swift Testing: 2 tests passed
• Python evaluator tests: 6 passed
• bash scripts/ci-basic-checks.sh: passed
• staged diff and conflict-marker checks: passed

Research

See docs/superpowers/specs/2026-07-30-voice-quality-research.md for source evaluation, licensing notes, phased recommendations, and the proposed real-corpus benchmark.

[IchenDEV]

Summary

• improve Apple Speech, WhisperKit, and Qwen recognition paths with contextual vocabulary, safer model selection, decoding fallbacks, VAD, and high-quality 16 kHz mono preprocessing
• make LLM correction evidence-bounded through request snapshots, faithful prompts, truncation handling, protected-token checks, semantic fidelity checks, and safe fallback to prepared ASR text
• improve dictionary replacement behavior, deferred replacement safety, and OCR-backed screen context
• add regression coverage plus a dependency-free CER/WER, terminology, entity-fidelity, hallucination, and latency evaluator
• document the research findings, open-source references, rollout plan, and remaining accuracy gaps

Why

Recognition quality was being reduced by inconsistent language and dictionary context, brittle Whisper model resolution, unsuitable Qwen input audio, and overly shallow decoding fallbacks. The correction layer could also accept truncated or semantically altered LLM output, while the project lacked a repeatable way to measure raw ASR quality separately from post-processing quality.

Impact

Users should get more robust recognition across Apple Speech, Whisper, and Qwen paths, with better handling of personal vocabulary and fewer correction-stage changes to numbers, dates, units, URLs, email addresses, file paths, quoted text, polarity, and self-corrections. When correction cannot be proven faithful, OpenType now preserves the prepared recognition result instead of inserting a risky rewrite.

The included example corpus validates the evaluator pipeline only. It is not evidence of production accuracy improvement; real-device recordings are still required for comparative benchmarking.

Validation

• swift test: 529 XCTest tests passed, 7 environment-dependent tests skipped
• Swift Testing: 2 tests passed
• Python evaluator tests: 6 passed
• bash scripts/ci-basic-checks.sh: passed
• staged diff and conflict-marker checks: passed

Research

See docs/superpowers/specs/2026-07-30-voice-quality-research.md for source evaluation, licensing notes, phased recommendations, and the proposed real-corpus benchmark.

[IchenDEV]

Summary

• improve Apple Speech, WhisperKit, and Qwen recognition paths with contextual vocabulary, safer model selection, decoding fallbacks, VAD, and high-quality 16 kHz mono preprocessing
• make LLM correction evidence-bounded through request snapshots, faithful prompts, truncation handling, protected-token checks, semantic fidelity checks, and safe fallback to prepared ASR text
• improve dictionary replacement behavior, deferred replacement safety, and OCR-backed screen context
• add regression coverage plus a dependency-free CER/WER, terminology, entity-fidelity, hallucination, and latency evaluator
• document the research findings, open-source references, rollout plan, and remaining accuracy gaps

Why

Recognition quality was being reduced by inconsistent language and dictionary context, brittle Whisper model resolution, unsuitable Qwen input audio, and overly shallow decoding fallbacks. The correction layer could also accept truncated or semantically altered LLM output, while the project lacked a repeatable way to measure raw ASR quality separately from post-processing quality.

Impact

Users should get more robust recognition across Apple Speech, Whisper, and Qwen paths, with better handling of personal vocabulary and fewer correction-stage changes to numbers, dates, units, URLs, email addresses, file paths, quoted text, polarity, and self-corrections. When correction cannot be proven faithful, OpenType now preserves the prepared recognition result instead of inserting a risky rewrite.

The included example corpus validates the evaluator pipeline only. It is not evidence of production accuracy improvement; real-device recordings are still required for comparative benchmarking.

Validation

• swift test: 529 XCTest tests passed, 7 environment-dependent tests skipped
• Swift Testing: 2 tests passed
• Python evaluator tests: 6 passed
• bash scripts/ci-basic-checks.sh: passed
• staged diff and conflict-marker checks: passed

Research

See docs/superpowers/specs/2026-07-30-voice-quality-research.md for source evaluation, licensing notes, phased recommendations, and the proposed real-corpus benchmark.