Linnet linnet-en-35m v1

Try it: the demo Space runs these weights on a CPU with the linnet-tts wheel.

Weights of the Linnet on-device text-to-speech model (41 M parameters, English, 24 kHz, streaming). This repository is a mirror of the GitHub release weights-v1 of https://github.com/VoxRT/LinnetTTS; the files are byte-identical and SHA256SUMS lists them.

file	purpose
linnet-en-35m.fp16.safetensors	runtime weights (what the Rust runtime and every binding load)
linnet-en-35m.safetensors	fp32 reference weights (PyTorch package, golden tests)
tokenizer.model	SentencePiece model
voices/linnet-en-f1.npy, voices/linnet-en-m1.npy	built-in voices (float32 [T, 48])
MANIFEST.json, SHA256SUMS	contents, checksums, reference synthesis md5

Code and bindings (Linux C, Python wheel linnet-tts, Android AAR, iOS Swift package): https://github.com/VoxRT/LinnetTTS (Apache-2.0). The weights are licensed under the Linnet Weights License (WEIGHTS_LICENSE.md), which is not an open-source license: it permits free use of the unmodified weights with the built-in voices inside your own products, including commercial and closed-source ones, and reserves custom voices, modification of the weights, model training and general-purpose text-to-speech services to separate written licenses from VoxRT (help@voxrt.com). The license text governs; this summary does not. The model card follows.

Examples

Default parameters (temp 0.3, cfg 2.0, seed 0), fp16 weights, Rust runtime. The model has no text normalisation: simple digit forms are usually read correctly, dense raw text (currency amounts, percentages, slash or ISO dates, phone numbers) should be converted to words before synthesis.

voice	text	audio
linnet-en-f1	Good morning, and welcome back. The meeting starts at nine, and the coffee is already on the table.
linnet-en-f1	There are 24 students in room 12, and the class starts at 9:30.
linnet-en-m1	Good morning, and welcome back. The meeting starts at nine, and the coffee is already on the table.
linnet-en-m1	There are 24 students in room 12, and the class starts at 9:30.

Model card: linnet-en-35m

Summary

Linnet is a lightweight on-device text-to-speech model: an autoregressive flow language model that emits one continuous 48-dimensional latent per 80 ms step (two 24-dimensional codec frames at 25 Hz) with a single flow evaluation per step, followed by a Vocos-style neural codec decoder that turns latents into 24 kHz audio. The model speaks English in a voice given as a voice-prompt latent, streams audio from the first step, and runs in real time on a single phone- or desktop-class CPU core. It ships with two built-in voices; the encoder that produces latents from recordings is not part of this release, and creating new voices requires a separate license (see WEIGHTS_LICENSE.md).

Parameters	41.4 M total: flow LM 31.5 M (text table 3.1 M), codec decoder 9.9 M
Sample rate	24 kHz, mono
Latent grid	25 Hz codec frames (24-dim); LM step 12.5 Hz (48-dim)
Context	voice prompt (the built-in voices: 6 s → 76 LM positions; training used ≈5 s prompts) + text tokens + generated steps, capped by max_frames (default 300 steps = 24 s)
Text frontend	SentencePiece unigram, 6000 pieces, no phonemizer, no text normalisation inside the model
Sampling	1 flow step per frame; temp scales the Gaussian draw (default 0.3); classifier-free guidance cfg on the trunk output (default 2.0); EOS head
Weights	safetensors linnet-tts/1; fp16 matrices for the runtime, fp32 reference
Languages	English only

Architecture

• Trunk: 6 pre-LayerNorm transformer layers, width 512, 8 heads, rotary positions (interleaved pairs, max period 10⁴), GELU (tanh) FFN with 4× expansion and no biases. Inputs: a learned "begin voice" embedding, the projected voice-prompt latents, the text token embeddings, then the projected previously generated latents. KV cache for streaming.
• Flow head: 6 AdaLN MLP blocks of width 512 conditioned on the trunk state and on two sinusoidal time embeddings; at inference the head is evaluated once per step on x0 ~ N(0, temp·I) and the latent is x0 + v. Latents are generated in a normalised space and de-normalised with stored mean/std.
• EOS head: a single logit on the trunk state; the step on which it fires is not emitted.
• Codec decoder: 1×1 projection 24→384, 2 causal ConvNeXt blocks at 25 Hz, ×4 upsampling to the 100 Hz STFT-column grid, 8 causal ConvNeXt blocks, a linear head producing 513 log-magnitudes and 513 phases per column, inverse FFT (1024) with a 960-sample Hann window, hop 240, causal overlap-add. The decoder reproduces its input delayed by 720 samples; the runtime compensates for this.

Every operator is causal; there is no lookahead anywhere in the chain. See docs/streaming-contract.md and docs/weights-format.md.

Training data and licenses

All training data comes from public corpora whose licenses permit commercial use. No proprietary or user data was used. Common Voice was used for validation only. Each corpus was resampled to 24 kHz, filtered for bandwidth and noise, and cut into utterances; the recordings themselves are not redistributed with this model. The copyright of the recordings remains with their original creators.

Codec (encoder + decoder, trained first and then frozen; the decoder is part of this release): 5 612 hours in 8 languages.

corpus	language	hours	license	source
CML-TTS (de, es, fr, it, pt)	de/es/fr/it/pt	2 302	CC BY 4.0	OpenSLR 146, LibriVox recordings
HiFiTTS-2	en	1 220	CC BY 4.0 (manifests); audio from LibriVox (public domain)	nvidia/hifitts-2
Emilia-YODAS (de, fr, zh)	de/fr/zh	956	CC BY 4.0 (the base Emilia set, CC BY-NC 4.0, was not used)	amphion/Emilia-Dataset, derived from YODAS2 (YouTube, CC BY 3.0)
LibriTTS-R	en	507	CC BY 4.0	OpenSLR 141
HUI-Audio-Corpus-German (clean)	de	248	CC0 (as stated by the authors, arXiv 2106.06309); LibriVox recordings	opendata.iisys.de
YO-CPT-ru	ru	163	CC BY 4.0 (annotations); audio CC BY 3.0 (YouTube, via YODAS2)	NCSpeech/YO-CPT-ru
AISHELL-3	zh	82	Apache-2.0	OpenSLR 93
Thorsten-Voice TV-44kHz-Full	de	39	CC0 1.0	Thorsten-Voice/TV-44kHz-Full
Crowdsourced Latin American Spanish	es	36	CC BY-SA 4.0	OpenSLR 61, 71–75
Vibravox (speech_clean, headset)	fr	31	CC BY 4.0	Cnam-LMSSC/vibravox
SIWIS French Speech Synthesis Database	fr	11	CC BY 4.0	Edinburgh DataShare
Other public speech data	ru	17	commercial use permitted	—

Flow LM (this release): 6 012 hours of English speech with word-level alignments.

corpus	hours	license	source
Emilia-YODAS (en)	4 192	CC BY 4.0	amphion/Emilia-Dataset
HiFiTTS-2	1 220	CC BY 4.0; audio from LibriVox (public domain)	nvidia/hifitts-2
LibriTTS-R	527 (LM-side selection; the codec used 507 h of the same corpus)	CC BY 4.0	OpenSLR 141
synthetic speech for numbers, dates, currencies and similar formats	74	generated with Qwen3-TTS (Apache-2.0, no restrictions on outputs), in the voices of corpus speakers	Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice

About 10 % of the LM utterances also carried an alternative written form of the transcript (digits instead of words), generated with a text model.

Notes on specific licenses:

• CC BY-SA 4.0 (Latin American Spanish, 0.6 % of the codec data): whether trained weights are an "adaptation" under ShareAlike is legally unsettled. The set is credited here; the weights are published under the Linnet weights license.
• YouTube-derived corpora (Emilia-YODAS, YO-CPT-ru): the CC BY licenses were declared by the video authors and collected by the corpus authors, who honour removal requests.

Training-time tools that are not part of the release: a WavLM model (MIT) served as a feature teacher for the codec encoder; the decoder follows the Vocos architecture (Vocos: Siuzdak, 2023) in our own implementation; the LM was trained with a training library whose code is not included here. Please cite the Emilia papers (arXiv 2407.05361, 2501.15907) when citing the data.

Evaluation

• Intelligibility: on held-out voices the generated speech is transcribed by an independent ASR model with word error rates of 15–30 % on short sentences, dominated by ASR tokenisation differences (contractions, "ad"/"add", "route"); canonical test sentences are reproduced verbatim.
• Listening: in blind comparisons against the codec's own reconstruction of real speech, listeners rank the model's output slightly rougher than the reconstruction in roughly half of the pairs; against real speech, the codec reconstruction itself is distinguishable. Overall quality was judged acceptable for release; the remaining roughness is a known limitation (below).
• fp16 vs fp32 weights: no audible difference in blind listening; the two produce different but equally valid samples because the one-step flow amplifies tiny numerical differences into a different trajectory.
• Numerical parity: the PyTorch package reproduces the training-side implementation bit for bit on the reference vectors in tests/goldens/ref; the Rust runtime matches the PyTorch package within 1e-4 on latents (measured ≤ 5e-5) and within 1e-2 on audio (SNR ≥ 70 dB on the goldens; the decoder amplifies latent-level differences).

Performance (Rust runtime, fp16 weights)

One step produces 80 ms of audio (one LM step plus two decoder frames). Output is bit-identical for any thread count. Real-time factor = mean step time / 80 ms; "first audio" is the prime pass plus the first step, measured in a command-line or test harness unless stated.

device	threads	per step	first audio after text	real-time factor
desktop x86-64, Core i9-13900K (AVX2)	1	7.0–7.4 ms	75–103 ms	9 %
desktop x86-64, Core i9-13900K (AVX2, auto)	4	2.5 ms	25 ms	3 %
Snapdragon 662, Cortex-A73 (Android)	1	42.4 ms	—	53 %
Snapdragon 662, 4 big cores (Android, auto)	4	14.3 ms	≈250 ms (sample app)	18 %
Raspberry Pi Zero 2 W, Cortex-A53 1 GHz (Linux)	1	158 ms	—	198 %
Raspberry Pi Zero 2 W, 4 cores (Linux, auto)	4	59–60 ms	—	75 %
iPhone 11, A13 (iOS)	1 (auto)	6.7–7.1 ms	107–116 ms	8–9 %
iPhone 13 Pro Max, A15 (iOS)	1 (auto)	5.5 ms	85 ms	7 %
iPhone 13 Pro Max, A15 (iOS)	2	4.6 ms	51 ms	6 %
Apple M5 (macOS)	1	3.3 ms	—	4 %

Measured without real-time pacing; when synthesis is paced to playback, mobile power management raises the step time 3–4× (see bindings/README.md, "Playback and step time"). An integer (W8A8) variant for weaker devices and a browser build are planned for a later release.

Compared with other CPU TTS models

Same desktop (Core i9-13900K, one performance core per thread), each model through its published Python package, three English texts, median of three runs. RTF = generation time / audio duration, including text processing and the prime pass. Speed only, on English text; quality is not measured here. Linnet is English only; Kokoro serves eight languages from the one model measured, pocket-tts and Piper ship one model per language (their English models are measured). Sizes are decimal megabytes. Method, versions and the reproduction script: docs/benchmarks.md.

model	languages	params	weights	cloning	streaming	RTF, 1 thread	RTF, 4 threads	first audio (short / long text)	peak RSS	same sentence
Linnet linnet-en-35m (fp16, Rust)	English	41 M	83 MB	yes	every 80 ms	10.6 %	3.4 %	87 / 133 ms	187 MB
pocket-tts 4.0.1 (PyTorch)	English (7 languages, one model each)	100 M	236 MB	yes	every 80 ms	23.8 % (2 cores)	24.1 %	71 / 82 ms	968 MB
Kokoro-82M (ONNX fp32)	8 languages in one model	81 M	354 MB	no	no	51.7 %	17.2 %	1.5 / 9.6 s	1246 MB
Piper en_US-ljspeech-medium (ONNX)	English (one model per voice, many languages)	16 M	64 MB	no	per sentence	5.8 %	2.1 %	180 / 283 ms	503 MB
Piper en_US-ljspeech-high (ONNX)	English (one model per voice, many languages)	28 M	114 MB	no	per sentence	42.5 %	12.4 %	1.4 / 2.1 s	639 MB

The audio column is "Good morning, and welcome back. The meeting starts at nine, and the coffee is already on the table." from each model with its default settings (Linnet: linnet-en-f1, temp 0.3, cfg 2.0, seed 0; pocket-tts: voice alba, CC BY 4.0, Kyutai; Kokoro: af_heart, Apache-2.0; Piper: LJSpeech voices, public-domain dataset, 22.05 kHz).

Built-in voices

Both built-in voices are synthetic: they were designed from a text description with Qwen3-TTS-12Hz-1.7B-VoiceDesign (Apache-2.0, no restrictions on outputs), so they are not clones of any real person. A 6 s utterance of each designed voice was resampled to 24 kHz, peak-normalised to −3 dBFS and encoded into the voice latent that ships here; no denoising was applied.

voice	file	description used for the design
linnet-en-f1	voices/linnet-en-f1.npy (76 × 48)	a bright, confident female voice in her late twenties, American accent, natural energy, news-presenter clarity
linnet-en-m1	voices/linnet-en-m1.npy (76 × 48)	a deep, intense male voice with an American accent, documentary narrator, calm and authoritative

The voices may be used as product voices under WEIGHTS_LICENSE.md. Because they are synthetic, no voice-performer credit applies.

Intended use and limitations

• English only; other languages produce garbage.
• Numbers, dates, currencies, phone numbers, version strings: the model reads written-out words reliably, raw digit formats only sometimes (for example "1359" or "$19.99" are often misread). Normalise text to words before synthesis.
• Voice latents: the model is conditioned on a latent of about 5–6 s of clean speech from one speaker, peak-normalised to −3 dBFS (the built-in voices: 6 s); short or spontaneous prompts drift in timbre, and room acoustics or noise degrade the voice. This describes the model's behaviour, not a permission: the standard license allows only the built-in voices and voices VoxRT has expressly authorised in writing.
• Audio quality: a slight roughness is audible on headphones in some utterances; not audible on laptop speakers.
• Under the standard license the model may be used with the built-in voices and with voices VoxRT has expressly authorised in writing. Cloning any voice (including your own, with consent), designing new voices, fine-tuning the weights or saving modified, quantised or converted copies of them, training models with the weights or their output, and offering general-purpose text-to-speech to others as an API or service, free or paid, require a separate written license (help@voxrt.com). It may not be used to impersonate a real person deceptively or to present its output as a genuine recording. See WEIGHTS_LICENSE.md.
• Not designed or certified as the sole means of conveying safety-critical or emergency information; see WEIGHTS_LICENSE.md, Section 10.

Versioning

linnet-en-35m v1: trained in October 2026. The weights file records the training step and whether EMA weights are stored in its metadata.