mlboydaisuke commited on
Commit
f73ec1e
·
verified ·
1 Parent(s): 0a05218

README: the 512-position KV cache and what CoreAIKit does at its end; iPhone 18 Pro: the ios/ graphs did not load (2026-11-13)

Browse files
Files changed (1) hide show
  1. README.md +8 -0
README.md CHANGED
@@ -110,6 +110,10 @@ continuous-feedback path is quant-sensitive — same split mlx-community uses).
110
 
111
  ## iPhone 18 Pro, the JIT bundles in `ios/` (load only)
112
 
 
 
 
 
113
  Measured 2026-09-26 on an iPhone 18 Pro (iPhone19,2, iOS 27.0 build 24A437, h19p) with the zoo's
114
  DecideGate app in its load-only mode, without the increased-memory entitlement. Each first load was the
115
  first after a fresh install of the app. The call is one run on all-zero inputs; a graph with a KV state
@@ -139,6 +143,10 @@ let tts = try await VoxCPM2TTS(paths: .standard(artifactsRoot: root, lm: .int8))
139
  let wav = try await tts.synthesize("On device speech synthesis, running entirely on your iPhone.") // 48 kHz Float PCM
140
  ```
141
 
 
 
 
 
142
  ## Verification
143
 
144
  Reimplemented in exportable Core AI overlays and gated end-to-end against the official model: backbone /
 
110
 
111
  ## iPhone 18 Pro, the JIT bundles in `ios/` (load only)
112
 
113
+ On an iPhone 18 Pro with iOS 27.2 beta (build 24B5099f), a test app ended twice on 2026-11-13 while CoreAIKit's `VoxCPM2TTS` loaded these `ios/` graphs, so VoxCPM2 has not produced audio on that phone. One run left a crash report: an abort from `MTLReportFailure` in `-[MPSGraphExecutable serializeToMPSGraphPackageAtURL:descriptor:]`, during the phone's on-device compile, with no assertion message. Both runs had loaded VoxCPM-0.5B and VibeVoice in the same process first; VoxCPM2 alone was not tried, and the cause is not known. The `macos/` graphs are the same seven files as `ios/` (the same sha256), and on a Mac (M4 Max, macOS 27.0) `VoxCPM2TTS` loaded them and produced audio on 2026-11-14.
114
+
115
+ For speech on an iPhone 18 Pro, use [VoxCPM-0.5B-CoreAI](https://huggingface.co/mlboydaisuke/VoxCPM-0.5B-CoreAI) (16 kHz): it loaded and produced audio in the same runs.
116
+
117
  Measured 2026-09-26 on an iPhone 18 Pro (iPhone19,2, iOS 27.0 build 24A437, h19p) with the zoo's
118
  DecideGate app in its load-only mode, without the increased-memory entitlement. Each first load was the
119
  first after a fresh install of the app. The call is one run on all-zero inputs; a graph with a KV state
 
143
  let wav = try await tts.synthesize("On device speech synthesis, running entirely on your iPhone.") // 48 kHz Float PCM
144
  ```
145
 
146
+ A clip's text and speech share the base and residual decode graphs' fixed KV cache of 512 positions: one position per text token and one per speech frame. A clip is at most 300 frames (48 s) by default (`maxFrames`), so a text of up to 211 tokens (about 800 characters of English) always fits, and a longer one can run out of room. From the CoreAIKit release after 0.7.3, the clip then ends where the cache ends, before the end of the text, and a text that leaves no room for speech throws `VoxCPMError.textTooLong` (VoxCPM2 shares VoxCPM's errors). CoreAIKit 0.7.3 and earlier do not stop at the end of the cache, and in a Mac test the kit raised no error and kept producing audio past it, which has not been listened to.
147
+
148
+ `KitSpeaker` and `CoreAI.speak` pass the whole text at once: split a long text into sentences and synthesize them one by one.
149
+
150
  ## Verification
151
 
152
  Reimplemented in exportable Core AI overlays and gated end-to-end against the official model: backbone /