coder543/paradee-8m-glade

Paradee-8M for Glade

Model assets and runtime configuration for Paradee-8M v1.1 in Glade. This single-voice American English distillation of Kokoro accepts ordinary text or explicit phonemes and produces mono 24 kHz PCM with its trained Heart voice.

The default v1.1 checkpoint contains 7,884,450 learned parameters: 4,231,828 in the text side and 3,652,622 in the decoder, including the 990,018-parameter waveform head. Counts come from the original modules before materializing weight normalization. Fixed signal-processing tables and the bundled text frontend are counted separately.

Packages

Both releases preserve the upstream sentence-first policy and the full 510 phonemes plus two boundary tokens. Each sentence is phonemized in context; oversized sentences split at the last fitting phoneme-space. The app catalog does not offer a short-chunk profile.

Location	Release	Waveform architecture	Download
Root	1.1	Source-filter head	22.2 MB
v1.0/	1.0	Upsampling generator, ISTFT and phase correction	23.0 MB

V1.0 contains 8,080,406 learned parameters. Its text/prosody/acoustic stages use INT8 weights and FP16 activations; the GPU vocoder uses FP16 weights with FP32 normalization. It requests ANE for text and GPU for the other three stages. Its original native phase filter and waveform reconstruction are retained. Both packages include the 11.0 MB American English frontend, FP16 CPU parameter storage with FP32 recurrence, and 2,048 acoustic frames. Select one complete package from a single repository snapshot.

V1.1 text and execution

The default 512/2,048 package is 22.2 MB, including the 11.0 MB American English Misaki frontend: trained spaCy tokenization/POS, contextual lexical processing, number handling and the original BART pronunciation fallback. Text input works directly through Glade without a separate frontend download or Python runtime.

Four Core AI assets implement the text network, pitch/noise prediction, acoustic trunk and frame-rate waveform head. Their INT8 weights use FP16 activations, with ANE text/head and GPU prosody/acoustic in the default 512-token/2,048-frame profile. CPU parameters use FP16 storage and FP32 recurrence. This hybrid configuration avoids failed ANE specialization attempts for the larger prosody/acoustic assets observed on the M3 Mac.

V1.1's waveform head predicts harmonic gains, phase adjustments, noise gains and pulse shifts at 80 frames per second. Native Swift reconstructs the original 64-harmonic source, pitch-modulated white noise and short-window burst noise. Its centered, reflect-padded STFT grids use 600-point/150-sample and 150-point/75-sample transforms. Fundamental phase accumulates in FP64 before FP32 interpolation. Exact frequency-warp matrices and Hann windows contain 26,654 FP32 values, stored in signal.f32 with the layout in signal.json.

Schema 2 metadata.json specifies 512 token slots, including boundaries, and 2,048 acoustic frames, along with the masking and signal conventions. Four .aimodel assets and cpu.f16/cpu.json supply the learned stages. Padding is excluded from utterance normalization and masked between convolutions; waveform reconstruction consumes only valid head rows. Text follows the sentence-first policy above. A duration overflow retries smaller parts before PCM delivery, preserving the complete input. Device specialization occurs on first use; compiled device caches are not included.

Qualification

Both full-context packages use upstream sentence-first planning: full sentences up to 510 phonemes plus two boundary tokens, with oversized sentences divided at the last fitting phoneme-space. The following Release results are from an M3 MacBook Air with 16 GB, macOS 27.0.1 and Swift 6.4. They are medians of three warm ordinary-text passes, including Misaki, planning and full waveform synthesis. Initial preparation, warmup and WAV writing are separate. The voice is Heart, speed 1 and seed 42, with one synthesis lane and complete delivery.

Release	Input	Generated audio	Median synthesis	RTFx
v1.0	Sunlight, three short sentences	6.900 s	494.57 ms	13.95×
v1.1	Sunlight, three short sentences	6.900 s	121.29 ms	56.89×
v1.0	Complete JFK reference transcript	748.750 s	13.794 s	54.28×
v1.1	Complete JFK reference transcript	748.750 s	8.276 s	90.47×

RTFx uses actual generated audio duration. Both releases generate 12 min 28.75 s for the complete JFK text. These matched planning and capacity measurements do not establish a perceptual quality ranking.

Bounded M3 Instruments captures of both exact packages measured five calls per neural stage:

Release	Stage	Calls	ANE predictions	GPU intervals
v1.0	Text	5	5	0
v1.0	Prosody	5	0	23
v1.0	Acoustic	5	0	29
v1.0	Vocoder	5	0	66
v1.1	Text	5	5	0
v1.1	Prosody	5	0	23
v1.1	Acoustic	5	0	25
v1.1	Head	5	5	0

These intervals verify execution placement, not arithmetic utilization.

New complete JFK waveforms from these exact packages were transcribed by Cohere Transcribe revision 76b8b23e8607f35f0265a23d481b338fb0e26aea, with Whisper English normalization:

Release	Word errors / normalized reference words	WER
v1.0	16 / 2,220	0.72%
v1.1	15 / 2,220	0.68%

These checks confirm coherent full-text coverage and include recognition errors; they are not perceptual or voice-similarity scores. Full-pipeline peak memory has not been measured. The older short-context and capacity-first measurements use different planning or profiles and are not measurements of these packages. Current iPhone and Watch qualification is pending.

Attribution

V1.1 weights: 330617ad1f097b4f6b02c103e54b39006c090b45. V1.1 implementation: 606e16e12f53bc6dfb814fe89fef338726687161. V1.0 weights: f662642d44c03c17588e4176469c54d462c0b623. V1.0 implementation: 9c8b4d7504cbee7e64de2d0341bb690f0b0ab708. The original model/code and Misaki use Apache-2.0; trained frontend components retain their bundled notices. No additional training is performed. Download weights and configuration from one repository snapshot.