Repository navigation
Releases: mrlm-net/voice-goio
Releases · mrlm-net/voice-goio
Release list
v0.16.2
v0.16.1
v0.16.0
Added
speaker.Speaker.SetGain(ch, gain)andGain(ch): a volume per channel (radio, intercom, PA), 0 muted to 1 full, clamped. It applies at once to what is playing on the channel, chimes included (the next part sent to the device, about 0.1 s on Windows), and to the channel's player when it opens later. For a cockpit audio panel: the MyCrew app sets each channel from the aircraft's volume knob and receive key.audio.Player.SetGainandGain: the player's volume, applied as the audio goes to the device; a session recording keeps the audio as given.
v0.15.0
v0.14.0
v0.12.1
Fixed
- speaker: one voice says one line at a time across channels. A captain talking to the cabin on the intercom no longer makes a PA at the same moment: a line in the same voice (same model and speaker) waits until that voice's current line ends. Different voices can still overlap, and chimes stay in order on their channel.
v0.12.0
Added
- speaker: an output device per channel.
SetDeviceFor(ChannelRadio | ChannelIntercom | ChannelPA, id)routes one channel; "" goes back to the main device, which stays the default.State().DeviceForlists the routed channels. - speaker: a cabin PA channel.
SayPA(text, voice)speaks on its own queue and player, independent of the intercom, so a PA and an intercom call can overlap.Chime(ChimePA)plays on the PA, andCliprenders a PA through the cabin chain. - speaker:
PAChain, a light cabin-speaker sound: band-limited to about 300 Hz–4 kHz with two early reflections.
v0.11.2
Added
- speaker: cabin chimes on the intercom, made in code with no files.
Chime(ChimeCall | ChimePA | ChimeSeatbelt)queues one in order withSayIntercom: the two-tone call, a soft ding before a PA, or the seat-belt bong. After a call chime, the next intercom item waits 1.5–3 s for the pickup.Cliprenders a chime, andChimePCM/ChimeRateexport its sound. A chime needs no voice model.
v0.11.1
Fixed
- tts/piper: a text of several sentences is said whole. Before, only the first sentence came back: piper writes each sentence's audio as soon as it is computed, and the pause while it computes the next one is longer than the idle gap. The text is now said one sentence at a time, joined with piper's 0.2 s sentence silence (#22).
v0.11.0
Install helpers for apps that download their voices on first start.
voices.InstallPack(ctx, pack, dir, progress): downloads a pack's models and checks each against its recorded SHA-256. It resumes interrupted downloads and skips models that are already installed and valid.Manifest.PackSize(pack): the download size of a pack. Current sizes: core 714 MB, en 1031 MB, all 1542 MB.tts/piper.Install(ctx, dir, progress): downloads piper 2023.11.14-2 for Windows (rhasspy/piper, MIT). It checks the zip against a pinned SHA-256 (f3c58906…efcea) and unzips it todir/piper, the layout the speaker uses by default.