coder543/pocket-tts-glade

Pocket TTS for Glade

Model assets and runtime configuration for Kyutai Pocket TTS in Glade. Ordinary text produces mono 24 kHz PCM, with complete and incremental synthesis interfaces. The original multilingual profiles and six smaller English profiles are independent downloads.

Small English models

The October 2026 English checkpoints are available in the directories below. Each base model also has Kyutai's reward-finetuned REFL variant. All six retain Alba, Anna, Reader A, Reader B, Kokoro Heart and Kokoro Adam conditioning, plus the native encoder for reference-audio conditioning.

Model	Repository directory	Complete download	M3 CPU RTFx	M3 ANE-preferred RTFx
1.9M	small/2m/	26.3 MB	70.36×	29.92×
1.9M REFL	small/2m-refl/	26.3 MB	70.19×	21.87×
6.9M	small/7m/	38.9 MB	39.56×	18.80×
6.9M REFL	small/7m-refl/	38.9 MB	34.30×	21.21×
18.2M	small/18m/	70.1 MB	44.79×	16.71×
18.2M REFL	small/18m-refl/	70.1 MB	44.56×	14.91×

Download sizes include all six voice prefixes and the separate reference-audio encoder. The advertised model parameter counts cover speech generation and waveform decoding; the reference encoder adds approximately 9.8M parameters.

These are warm Release text-to-audio medians of three passes on an M3 MacBook Air, 16 GB, macOS 27.0.1, using Alba, seed 42 and temperature 0.3. Input: “Sunlight pours through the trees. A gentle breeze moves across the clearing. Welcome to Glade.” Generated audio is 5.12–5.52 seconds, depending on checkpoint. The clock includes text preparation, language generation, waveform decoding and host state work; model preparation, warmup and WAV writing are separate. Explicit CPU execution is faster for these small models on this M3. These short measurements do not replace long-form or phone measurements of the new profiles.

Each small bundle has three FP16 .aimodel assets: language/flow, Mimi decoder and the reference encoder. It retains the native 1,024-position language cache, 50-position Mimi attention context, per-checkpoint Unigram tokenizer, text preprocessing, EOS threshold and one-step drifting or pointwise sampler. Codec entrypoints consume 1/2/4/8 consecutive latents while preserving state. The reference encoder loads only when audio conditioning is requested.

Bounded captures of all six ANE-preferred profiles show one ANE prediction per observed language/codec call and no target GPU intervals. Streaming and complete modes produce exactly matching PCM in the selected six-checkpoint controls. First 640 ms packets were available after 14–30 ms in warm native streaming checks; this measures generation, not the client's playback buffer. Recognition checks of selected samples confirm coherent words; perceptual scores and phone/Watch qualification of the new small profiles are not claimed.

Small checkpoint, tokenizer and preset revision: a72dba0d1a5f1387c104f9d3e17abb8ab60b4fae in kyutai/pocket-tts (CC BY 4.0). Original implementation: da629e60b68fc141a790daa05630735966b68720 (MIT). Per-variant notices retain the preset and example-voice attribution.

Original multilingual profiles

Languages are independent downloads. languages.json declares exact file groups, shared files and sizes; metadata.json declares the profile inventory. Each profile includes its original language weights, Mimi codec, tokenizer, text preparation rules and official preset conditioning. Each profile also includes Reader A, Reader B, Kokoro Heart and Kokoro Adam, precomputed from CC BY 4.0 LibriSpeech recordings and newly generated Kokoro examples supplied with GladeDemo. Pocket always uses voice conditioning; selecting no voice means the default preset. English is the default.

Language	Official presets	Download
English	Alba, Anna	153.8 MB
French	Estelle	150.7 MB
German	Juergen	150.0 MB
Portuguese	Rafael	150.0 MB
Italian	Giovanni	149.0 MB
Spanish	Lola	149.6 MB
Dutch	Daan	148.9 MB

All languages together total approximately 1.1 GB. Adding a language does not require downloading the other profiles. Glade uses SwiftModelHub to verify files against Hugging Face's native checksums and reuse unchanged files during updates.

Original multilingual execution

Each original language profile contains three source .aimodel assets: W8A16 language inference, an FP16 Mimi audio decoder and an FP16 reference-audio encoder. All request ANE; the reference encoder loads only for audio conditioning. The language graph retains six causal Transformer layers and the original one-step LSD flow head, producing an 80 ms latent at each autoregressive step. The full-history language cache has 1,024 positions, including the selected voice's prefix; prefill uses 32 positions.

The codec supports 1/2/4/8 consecutive latents with shared weights and preserves its native 250-position causal attention context and convolution state. Independent utterances are processed sequentially. Original tokenization, character replacements, sentence splitting, 50-token target, EOS conventions and short fade-in are retained. Unsupported excess capacity reports an error. No compiler specialization caches or reference recordings are included.

Current multilingual qualification

The schema-3 packages include native audio-reference conditioning, a 1,024-position language cache and the current matched tokenizer/preset assets. FP32 graph controls cover all seven languages, including the reference encoder and its streaming histories. The native text frontend is checked against 70 original-source cases across all seven languages.

Fresh warm M3 Release text-to-audio checks use the same three-sentence Sunlight sample in each language, one synthesis lane, temperature 0.3 and seed 42:

Language · preset	Generated audio	Median synthesis	RTFx
English · Alba	5.12 s	379.03 ms	13.51×
French · Estelle	5.20 s	308.56 ms	16.85×

The English result is a three-pass median; the French check is a single warm pass. Preparation and WAV writing are separate. Bounded English/French captures confirm ANE language/codec execution with no target GPU intervals. Selected Cohere checks confirm coherent words. General perceptual and voice-similarity scores are not claimed.

Earlier iPhone numbers used the prior 512-position packages. The updated schema-3 packages have Mac qualification; current phone and Watch qualification is pending. Short samples do not establish full-transcript throughput.

Attribution

Weights, tokenizers and official preset states: Kyutai, a72dba0d1a5f1387c104f9d3e17abb8ab60b4fae in kyutai/pocket-tts, CC-BY-4.0. Original implementation: Kyutai, da629e60b68fc141a790daa05630735966b68720, MIT. Licenses and preset attribution are included. Conversion and quantization change inference representation, not authorship or ownership.

M5 Max performance

MacBook Pro with M5 Max (18-core CPU, 40-core GPU, 64 GB), macOS 27.2.1, High Power mode; measured November 9, 2026. Complete text-to-audio synthesis, Alba, seed 42, temperature 0.3, one lane and three warm passes after one full warmup, using the same three-sentence Sunlight text above.

The original English profile reached 28.9× RTFx / 5,083 WPM with ANE preferred.

Small English model	CPU RTFx	CPU WPM	ANE-preferred RTFx	ANE-preferred WPM
18.2M	70.3×	11,803	29.2×	4,910
18.2M REFL	69.7×	11,707	29.2×	4,897
6.9M	48.3×	7,985	49.6×	8,209
6.9M REFL	48.5×	7,688	49.1×	8,007
1.9M	99.2×	17,432	68.8×	12,102
1.9M REFL	99.5×	17,487	70.5×	12,385

The clock includes the native text frontend, language generation, waveform codec and host work; preparation, warmup and WAV writing are excluded. WPM is 15 source words divided by generation time, multiplied by 60; it measures text throughput rather than speaking rate. Generated audio lasts 5.12–5.68 seconds: FP16 rounding and stochastic EOS can change duration across backends, so these are complete-synthesis rates, not isolated kernel speedups. ANE preference is a request; no new placement trace or transcript-quality qualification was collected for these timings. CPU remains clearly faster for the 1.9M and 18.2M models here; the 6.9M policies are close.