Talktastic v0.8.5

What's new in v0.8.5

• Fixed: converting an audio file with an RVC model URL did nothing. Running something like say -i before.wav -o after.wav --rvc "https://huggingface.co/..." would download (or find the cached) model, print Cached: ..., and quietly exit without converting anything. Passing --rvc a URL with no text is supposed to mean "just download this model," but Talktastic forgot to check whether you'd also given it an input file. It now converts the file as expected. Using a cached model's name instead of a URL was never affected.

• Regression test. A new integration test runs a file through -i with an RVC URL and checks that it actually writes the output, so this one stays fixed.

Asset

say.exe (Talktastic v0.8.5, 33.5 MB, win-x64, NativeAOT, .NET 10).

Commits since v0.8.4

• 705147b Fix --rvc with -i only caching the model instead of converting- e4a632f Bump version to 0.8.5

Talktastic v0.8.4

What's new in v0.8.4

• Smaller binary. say.exe dropped from 37 MB to 33.5 MB through NativeAOT size optimizations and switching all embedded resources (native DLLs, ONNX skeletons, manifests) to Brotli compression (quality 11).

• Smarter downloads. New FileDownloader with a resolver chain handles URL resolution, redirect following, and progress reporting in one pipeline. Downloads from voice-models.com, Google Drive, HuggingFace, GitHub Releases, and Weights.gg all go through the same path now.

• Streaming archive extraction. Downloaded models are extracted on-the-fly as they stream in -- no temp files, no double disk writes. Supports tar.gz, zip, gzip, brotli, and deflate.

• Download registry. Downloaded models are tracked in a local JSON registry so Talktastic knows what it already has and can skip re-downloading.

• --perf fix. Time-to-first-audio now shows the actual elapsed duration (e.g. 120ms) instead of just the source tag.

Asset

say.exe (Talktastic v0.8.4, 33.5 MB, win-x64, NativeAOT, .NET 10).

Commits since v0.8.3

• b98cd0b Add rebuild scripts, NativeAOT size optimizations, SDK 10.0.301 binary
• 869fe6b Unify download logic: ArchiveExtractor + multi-format archive support
• a068639 Remove dead Piper resolver code, migrate tests to ModelDownloader
• fce3c9d Add FileDownloader abstraction with resolver chain, registry, streaming extraction, and Brotli compression
• e47fccb Add comprehensive test coverage and XML documentation
• 7d814c0 Add integration tests for download error paths and Piper synthesis
• 3a0a0f0 Add ProgressTrackingReadStream and PthLoader edge case tests
• 31821b1 Switch all embedded resources to Brotli, fix --perf first-audio display

Talktastic v0.8.3

What's new in v0.8.3

• Windows N / KN support. Fixed a bug that prevented Talktastic from working on Windows N / KN editions (Neural voices still require the Windows Media Feature Pack to be installed, but Piper and SAPI will work without it now).

• RVC streaming. RVC voice conversion now starts playing as soon as the first chunk is ready instead of waiting for the whole utterance. Tunable via --rvc-chunk-size <seconds> (default 2.0) and --rvc-padding-length <seconds> (default 0.3).

• Performance and Instrumentation Improvements. On an iGPU/vram-constrained test machine, time-to-first-audio-play-back dropped from roughly 9 seconds to under 4 seconds. -- --perf argument added to give verbose instrumentation output. -- Model loads for Piper / RVC also run in parallel where applicable.

• Added support for more RVC model download sources. URLs from voice-models.com, Google Drive, Weights.gg, or HuggingFace now work directly with the --rvc parameter; Talktastic follows redirects, handles Google Drive's virus-scan interstitial, picks a sensible display name from the source page, and gives a clear error message when a CDN is unreachable.

• Automated audio quality validation. New integration tests round-trip synthesized audio back through a speech-to-text (STT) pipeline and additionally validate the zero-crossing rate to catch silent or garbled output that would otherwise pass a recall check. The same validator runs at runtime: if any RVC chunk comes back corrupt on GPU, it's silently re-run on a pre-warmed CPU session before reaching playback.

• DirectML stability fixes. Broader exception catch on init (now also handles BadImageFormat / TypeLoad / PlatformNotSupported); failure messages name the exception type for easier debugging, with graceful fallbacks to CPU.

Asset

say.exe (Talktastic v0.8.3's main console application, 37 MB, win-x64, NativeAOT, .NET 10).

Commits since v0.7.1

• 7f4c012 Sub-chunked streaming RVC + Settings refactor + coverage backfill
• b6ab76e Stream synthesized audio to the default device across all engines

v0.7.1

Fixed

• Voice resolution: Added substring-contains matching so partial names like Microsoft Aria correctly resolve before fuzzy matching can misfire on shared prefixes.

Changed

• URL-to-model mapping: voices.json and rvcs.json now contain actual JSON instead of tab-separated text. Reads old format automatically, always writes JSON.
• Renamed internals: All Registry references renamed to UrlMap to avoid confusion with the Windows Registry.
• README: Self-documenting example dialogue, folder URLs for Piper downloads, silent install example, type prefix disambiguation.

v0.7.0

Fixed

• Voice resolution: Microsoft Aria no longer misresolves to Microsoft Richard. Added substring-contains matching between exact and fuzzy match steps, so partial names like Microsoft Aria correctly match Microsoft Aria (Natural) before fuzzy matching can misfire on shared prefixes.
• Fuzzy match: Added LocalName to the fuzzy match chain for better fallback behavior.

Changed

• README: Rewrote examples with self-documenting dialogue, added punctuation, URL-first Piper flow.

Talktastic v0.6.0

Standalone Windows TTS CLI -- neural, SAPI, Piper voices with optional RVC voice conversion.

What's in the box:

• say.exe -- single NativeAOT binary, no runtime or installers needed
• Windows neural voices (Embedded Speech SDK)
• SAPI/WinRT legacy voices
• Piper ONNX voices (download by URL or shorthand, cached locally)
• RVC voice conversion (.onnx and .pth models, DirectML GPU acceleration)
• Output to speakers, .wav, .mp3, or .ogg (Opus)
• Import .wav/.mp3/.ogg for RVC-only processing
• SSML support, fuzzy voice matching, model management

See README for usage examples.