Talk Spotter 2.1.0-rc1 (experimental)

Release candidate built from the experimental-refactor branch. Not merged to main yet.

Highlights

• RTL-SDR DSP rebuilt (lib/dsp.py): polyphase decimation (~2x less CPU at 960 kS/s — matters on Raspberry Pi), proper instant-attack/slow-release AGC for SSB/AM replacing per-chunk peak normalization, fixed deviation-referenced scaling for FM, anti-alias filtering on the AM/FM paths, stateful DC blocking, and a comfort-noise dither that measurably improves Vosk recognition.
• Validated end-to-end against Vosk (real speech, modulated and demodulated, strong and weak signals): recognition ≥ 2.0.2 in all six scenarios; weak AM 5% → 79%, weak FM 88% → 94% F1.
• Thread-safety fixes: command parser locking (no more double-post risk), network posting moved off the audio thread, sources join worker threads before closing devices (fixes potential crash on exit).
• Dead code removed, POTA/SOTA posting deduplicated, 60+ new tests, new tools/validate_demod_e2e.py regression gate.

The .deb below is built by CI from this tag.

v2.0.2

What's New

• CI test gate: Release workflow now runs all unit tests before building .deb packages
• --version flag: python talk_spotter.py --version prints the current version
• Central version constant: lib/version.py is the single source of truth; SOTA User-Agent now stays in sync
• SOTA spotter tests: 18 new unit tests covering OAuth token storage/refresh and spot posting
• Updated CLAUDE.md: Reflects the lib/, sources/, spotters/, tests/ layout from 2.0.1

v2.0.1

Codebase Reorganization

Reorganized the project into a clean package structure:

• lib/ — core library modules (command_parser, grammar, transcription)
• spotters/ — network posting clients (dx_cluster, pota_spotter, sota_spotter)
• tests/ — all unit tests
• sources/ — audio source classes (unchanged)

Other changes

• Removed obsolete standalone scripts (kiwi_stream.py, rtl_stream.py), superseded by sources/
• Updated debian packaging for new directory layout

No functional changes — all 148 tests pass.

v2.0.0

What's New

• Add transceiver audio source with Hamlib CAT control via rigctld
• Add KiwiSDR auto-reconnect with exponential backoff (max_retries, retry_delay config options)
• Add unit tests for grammar, DX cluster, POTA spotter, and transceiver modules
• Fix duplicate words in Vosk grammar vocabulary
• Remove unused imports and deleted sound_card source

v1.1.3

What's new

• Local config overrides: Create config.local.yaml to override any settings without modifying the tracked config.yaml. Deep-merged at startup, gitignored by default.
• Default KiwiSDR host changed to ssi.proxy.kiwisdr.com so the developer's personal server isn't the out-of-box default.

v1.1.2

What's Changed

• Unit tests: 81 tests covering command parser happy paths, unhappy paths, edge cases, timeouts, and frequency parsing
• Vosk finalization fix: Force-flush Vosk when partial text is stable for 2 seconds — fixes grammar-constrained mode where Vosk recognizes words but never commits them as final results, which was causing "cancel" and other commands to be ignored
• Memory fixes: raw_text no longer grows unboundedly in idle state; audio buffer switched to bytearray for in-place operations
• Callsign slashes: "slash" and "stroke" now produce / in callsigns (e.g., VP2E/W1AW, W1AW/VP2/P)
• Removed keyword highlighting: Redundant feature from early development, now that the command parser handles all voice command recognition

v1.1.0

Voice command parser rewrite

The command parser has been significantly improved. The biggest user-facing change is that fields can now be spoken in any order — "frequency ... call ... end" works just as well as the original "call ... frequency ... end".

New features

• Any field order: callsign and frequency (and optional network) can be spoken in any sequence after the wake phrase
• "callsign" / "call sign": accepted as alternatives to "call"
• "cancel": say this at any point after the wake phrase to abort the command and return to idle without posting anything
• Compound number words: "twenty eight decimal five" now correctly parses as 28.5 MHz; previously tens+units combinations like "twenty eight" were misread as "208"
• Repeated keywords override: saying "call" (or any field keyword) a second time discards the earlier attempt and uses only what follows the second occurrence

Timeout improvements

• Idle timeout (10s): auto-finalizes after ~10 seconds of silence following the last spoken words, rather than counting from the wake phrase
• Session timeout (60s): hard cap from wake phrase as a safety net

Bug fixes

• NATO phonetic X-ray ("x-ray") now detected correctly in callsigns — Vosk splits it into two tokens ("x" and "ray") which are now merged before lookup

v1.0.1

Talk Spotter is now considered stable and out of beta. This is a quick patch release fixing one specific issue discovered after 1.0.0.

Fix

NATO phonetic X-ray was never detected. Vosk transcribes "x-ray" as two separate tokens (x and ray) rather than the hyphenated form. The grammar constraint also couldn't include "x-ray" due to the hyphen being invalid in Vosk grammar tokens. As a result, any callsign containing the letter X was silently dropped.

The fix adds "x ray" as a recognized two-word phrase in the Vosk grammar and merges x + ray bigrams into xray before the NATO alphabet lookup in the parser.

v1.0.0

Out of beta! This is the real one!

What's new

• Grammar-constrained recognition: Vosk is now constrained to command vocabulary (NATO phonetics, number words, keywords) by default, significantly reducing false positives from background radio noise. Use --no-grammar to disable.
• Spotting always enabled: The --spot-mode flag has been removed — spot posting is now on by default. Use --no-post to suppress posting for testing.
• "complete" terminator: Added "complete" as an alternative to "end" for finishing a voice command.
• Voice command parsing fixes:
  • Entire commands spoken in one breath now parse correctly (callsign no longer absorbs frequency words)
  • Premature timeouts from partial results being fed to the command parser are fixed
  • Frequency spoken in the same utterance as "end"/"complete" is now correctly captured

v0.2.0

What's New

• Add POTA spot posting support
• Add SOTA spot posting support with OAuth authentication
• Fix RTL-SDR freezing by separating read and process threads
• Add slow AGC to smooth audio level variations
• Add DC offset removal to FM demodulator
• Add second model path option to config
• Enhance error handling in SOTA authentication