awesome-open-tts-under-1b

All models, by parameter count

Inflect-Nano-v2

English, one fixed synthetic voice

4.0M

Apache 2.0 CPU yes

▶

Params: 3,966,721 (vendor-stated; no safetensors index)
License: Apache 2.0
Languages: English, one fixed synthetic voice
Streaming: No native streaming (punctuation-aware chunking)
TTFB / first-chunk: not published
CPU real-time: yes: RTF 0.0933, 10.72x real time (vendor, HF CPU Upgrade, 8 vCPU, 4 threads). Same repo reports RTF 0.6299 on a Ryzen 9 3900X
Quantized formats: not published
Integration path: DIY; HF only, no pip package

Open model ↗ Correct this row

Vaniq-Edge

English, one fixed voice

~ 9.0M

MIT * CPU yes

▶

Params: ~9.0M (derived from model.pth at fp32; card claims 8.5M)
License: MIT weights; required espeak-ng phonemizer is GPL v3
Languages: English, one fixed voice
Streaming: No native streaming
TTFB / first-chunk: not published
CPU real-time: yes: RTF 0.262 on 1 core (vendor, CPU unnamed, published only as text inside a chart image)
Quantized formats: not published (the card advertises an ONNX export that is absent from the repo tree)
Integration path: DIY; HF only, no GitHub repo

Open model ↗ Correct this row

Inflect-Micro-v2

English

9.4M

Apache 2.0 CPU yes

▶

Params: 9,356,513 (vendor-stated; no safetensors index)
License: Apache 2.0
Languages: English
Streaming: No native streaming
TTFB / first-chunk: not published
CPU real-time: yes: RTF 0.1593, 6.28x real time (vendor, HF CPU Upgrade, 8 vCPU, 4 threads)
Quantized formats: not published
Integration path: DIY; code. HF org is owensong, GitHub org is owenawsong

Open model ↗ Correct this row

Kitten TTS

English

~ 15.0M

Apache 2.0 CPU yes output stream

▶

Params: 15M to 80M
License: Apache 2.0
Languages: English
Streaming: Output-streaming (chunked)
TTFB / first-chunk: not published
CPU real-time: yes
Quantized formats: ONNX (ships as ONNX)
Integration path: DIY; ONNX runs in-browser and on edge

Open model ↗ Correct this row

Matcha-TTS

English (LJSpeech single-speaker, VCTK 108 speakers)

~ 18.2M

MIT CPU yes

▶

Params: 18.2M acoustic model, excludes vocoder (paper)
License: MIT
Languages: English (LJSpeech single-speaker, VCTK 108 speakers)
Streaming: No native streaming
TTFB / first-chunk: not published
CPU real-time: yes: RTF 0.411 (community, sherpa-onnx, Raspberry Pi 4 Model B Rev 1.5, 4 threads; RTF 0.941 at 1 thread)
Quantized formats: ONNX (in-repo exporter, prebuilt via sherpa-onnx)
Integration path: DIY; first-class sherpa-onnx support for CPU and embedded

Open model ↗ Correct this row

Matxa-TTS Catalan

Catalan, 4 accents, 8 speakers

~ 20.9M

GPL

▶

Restriction: Copyleft, and commercial deployment needs separate voice-artist licensing
Params: ~20.9M (derived at fp32; no exact count published)
License: GPL-3.0
Languages: Catalan, 4 accents, 8 speakers
Streaming: No native streaming (non-autoregressive OT-CFM)
TTFB / first-chunk: not published
CPU real-time: not published
Quantized formats: ONNX end-to-end (single file, includes the vocoder)
Integration path: ONNX e2e, or PyTorch with a separate vocoder plus compiled espeak-ng. A v2 line under BSC-LT is apache-2.0

Open model ↗ Correct this row

wfloat-tts

30.2M

MIT

▶

Params: 30,219,888 (safetensors.total)
License: MIT
Languages: not published
Streaming: not published
TTFB / first-chunk: not published
CPU real-time: not published
Quantized formats: not published
Integration path: First-party browser/WASM runtime (wfloat/wfloat-web) and React Native (wfloat/react-native-wfloat), both MIT

Open model ↗ Correct this row

Lahja

Arabic, Saudi dialect. Two voices across two repos

36.3M

No grant *

▶

Restriction: No license granted at all. No rights is worse than non-commercial
Params: 36,284,784 (safetensors.total)
License: none declared (cardData.license is null and the card states none)
Languages: Arabic, Saudi dialect. Two voices across two repos
Streaming: No native streaming
TTFB / first-chunk: not published
CPU real-time: not published
Quantized formats: not published
Integration path: Cleanest integration of any row here: plain Transformers VitsModel and AutoTokenizer, no external vocoder. The declared base repo returns 401, so lineage is not auditable

Open model ↗ Correct this row

Coqui VITS LJSpeech

English, single speaker, 22,050 Hz

36.3M

Apache 2.0 *

▶

Params: 36,321,264 (measured from the release checkpoint; no published count)
License: "apache 2.0" (string from TTS/.models.json); toolkit code MPL-2.0
Languages: English, single speaker, 22,050 Hz
Streaming: No native streaming
TTFB / first-chunk: not published
CPU real-time: not published
Quantized formats: not published
Integration path: pip install coqui-tts (maintained idiap fork). Functionally overlapping with Piper, which is also VITS-based and ships ONNX

Open model ↗ Correct this row

Ukrainian TTS

Ukrainian, 5 voices, automatic stress marking

40.7M

MIT

▶

Params: 40,685,869 generator (measured; the 446 MB checkpoint also carries 70,711,277 training-only discriminator params)
License: MIT
Languages: Ukrainian, 5 voices, automatic stress marking
Streaming: No native streaming
TTFB / first-chunk: not published
CPU real-time: not published
Quantized formats: not published
Integration path: pip install from GitHub. Released v6.0.0 is Tacotron2 + HiFi-GAN joint despite the repo name, frozen 2023-09-17; current work is on an unreleased branch

Open model ↗ Correct this row

Coqui YourTTS

3: en, fr-fr, pt-br. 38 released speaker embeddings, 16 kHz output

48.1M

Non-commercial

▶

Restriction: Non-commercial and no-derivatives (NoDerivatives arguably bars fine-tuning)
Params: 48,123,087 inference path (measured; the full state dict includes a training-only discriminator)
License: CC BY-NC-ND 4.0 weights, MPL-2.0 toolkit code
Languages: 3: en, fr-fr, pt-br. 38 released speaker embeddings, 16 kHz output
Streaming: No native streaming
TTFB / first-chunk: not published
CPU real-time: not published
Quantized formats: not published
Integration path: pip install coqui-tts (idiap fork). Weights frozen 2022-03-23; 16 kHz output is behind Kokoro and Piper

Open model ↗ Correct this row

Style-Bert-VITS2

Japanese (JP-Extra line); the older v2.x line is ja, zh, en

62.8M

AGPL

▶

Restriction: AGPL-3.0 network copyleft: serving a hosted voice agent triggers source disclosure
Params: 62,756,658 (measured); the mandatory JP-Extra DeBERTa front end adds 329,601,020
License: AGPL-3.0 code, CC-BY-SA-4.0 weights, LGPL-3.0 for the text and user dictionary
Languages: Japanese (JP-Extra line); the older v2.x line is ja, zh, en
Streaming: No native streaming (no chunk or yield path)
TTFB / first-chunk: not published
CPU real-time: not published
Quantized formats: ONNX export script (v2.7.0+); no GGUF or int8
Integration path: pip install style-bert-vits2; Gradio app and HTTP server. Code active to 2025-12-07, weights frozen 2024-02-05

Open model ↗ Correct this row

IMS-Toucan

471 supervised, ~7,200 zero-shot via language embeddings

81.6M

Apache 2.0

▶

Params: 81,563,686 full pipeline (measured: ToucanTTS 50,097,950 + vocoder 31,167,362 + embedding GAN 298,374)
License: Apache 2.0
Languages: 471 supervised, ~7,200 zero-shot via language embeddings
Streaming: No native streaming
TTFB / first-chunk: not published
CPU real-time: not published
Quantized formats: not published
Integration path: Clone and pip install, no package; needs espeak-ng. Zero-shot language lookups add ~3.75 GB beyond the weights

Open model ↗ Correct this row

Kokoro

8 languages, 54 voices

~ 82.0M

Apache 2.0 CPU yes output stream

▶

Params: 82M
License: Apache 2.0
Languages: 8 languages, 54 voices
Streaming: Output-streaming (chunked)
TTFB / first-chunk: not published (community reports sub-second full-utterance on CPU; hardware varies)
CPU real-time: yes
Quantized formats: ONNX community exports
Integration path: Pipecat service available; community LiveKit wrappers; verify plugin freshness

Open model ↗ Correct this row

Supertonic 3

31, plus a language-neutral mode

~ 99.0M

OpenRAIL CPU yes

▶

Restriction: OpenRAIL use restrictions that pass to every downstream user, and the project has announced archival
Params: ~99M (README prose only; ONNX assets, no safetensors index)
License: BigScience Open RAIL-M weights (cardData says the bare string "openrail"), MIT code
Languages: 31, plus a language-neutral mode
Streaming: No native streaming
TTFB / first-chunk: not published
CPU real-time: yes: average RTF 0.3 (vendor, Onyx Boox Go 6 e-reader, ARM CPU, airplane mode)
Quantized formats: ONNX fp32 only; community int8 and fp16 via supertonic-mnn
Integration path: Official Python SDK and serve command; ONNX Runtime examples in 11 languages. Community pipecat-supertonic plugin, explicitly unsupported. Archival announced 2026-07-23, Voice Builder retires 2026-08-31

Open model ↗ Correct this row

Kyutai Pocket TTS

6 (en, fr, de, pt, it, es; non-English as 24-layer preview variants)

~ 118M

CC-BY-4.0 CPU yes output stream TTFB gated

▶

Params: ~118M default checkpoint (derived at bf16; README says 100M; 24-layer language variants reach ~336M)
License: CC-BY-4.0 weights (per-voice terms vary, see kyutai/tts-voices), MIT code
Languages: 6 (en, fr, de, pt, it, es; non-English as 24-layer preview variants)
Streaming: Output-streaming (chunked)
TTFB / first-chunk: ~200 ms to first chunk (vendor, hardware unspecified; the MacBook Air M4 in the README attaches to the throughput claim, not this one)
CPU real-time: yes: ~6x real time (vendor, MacBook Air M4, 2 cores)
Quantized formats: Runtime int8 via torchao, no distributed artifact. Community ONNX, MLX, Candle, sherpa-onnx
Integration path: Pipecat first-party extra (pipecat-ai[pocket-tts]); own serve command; sherpa-onnx. No LiveKit plugin. HF repo is gated (auto): terms acceptance and a token required

Open model ↗ Correct this row

DhVaani-0.5

26 Indian languages + English

123M

Apache 2.0 gated

▶

Params: 122,798,800 (safetensors.total)
License: Apache 2.0 (inherited from the ZipVoice base per the card)
Languages: 26 Indian languages + English
Streaming: No native streaming
TTFB / first-chunk: not published
CPU real-time: not published
Quantized formats: not published
Integration path: DIY; ZipVoice derivative, vendors its own backend. Needs a 3 s to 10 s reference clip plus transcript. HF gated (auto)

Open model ↗ Correct this row

ZipVoice

Chinese, English

~ 123M

No grant * CPU no

▶

Restriction: Weights carry no license grant at all; commercial reuse is legally undefined
Params: 123M (README and paper, corroborated by the fp32 file size; no safetensors index)
License: Code Apache-2.0. Weights: no license key in cardData and none in the README frontmatter
Languages: Chinese, English
Streaming: No native streaming (punctuation chunking then batching)
TTFB / first-chunk: not published
CPU real-time: no: best case RTF 1.22, distilled at 4 NFE, single thread (vendor, Intel Xeon Platinum 8457C)
Quantized formats: ONNX fp32 and INT8 in-repo for the base and distill checkpoints
Integration path: Two first-party deployment paths: sherpa-onnx C++ on CPU, and an NVIDIA Triton runtime

Open model ↗ Correct this row

VieNeu-TTS v3 Turbo

Vietnamese, English, vi-en code-switching

131M

Apache 2.0 CPU yes output stream TTFB

▶

Params: 130,907,520 (safetensors.total)
License: Apache 2.0
Languages: Vietnamese, English, vi-en code-switching
Streaming: Output-streaming (frame-level, ONNX/CPU path only)
TTFB / first-chunk: ~300 ms to first audio (vendor, hardware unspecified)
CPU real-time: yes: RTF < 1, ~2-3x on a laptop (vendor, CPU unnamed)
Quantized formats: ONNX fp32 and ONNX INT8 in-repo (the CPU backbone runs int8 by default)
Integration path: DIY; vieneu on PyPI, FastAPI streaming demo. Hard dependency on the author's sea-g2p

Open model ↗ Correct this row

ArTST

Arabic (Classical, via ClArTTS)

144M

Non-commercial

▶

Restriction: Non-commercial
Params: 144,481,358 (derived at fp32; no safetensors index)
License: CC-BY-NC-4.0
Languages: Arabic (Classical, via ClArTTS)
Streaming: No native streaming
TTFB / first-chunk: not published
CPU real-time: not published
Quantized formats: not published
Integration path: Transformers, but needs speecht5_hifigan plus x-vector speaker embeddings. Heaviest integration in this table

Open model ↗ Correct this row

SILMA TTS v1

Arabic and English, zero-shot cloning

~ 150M

Apache 2.0

▶

Params: ~150M (vendor blog; config is the F5-TTS Small DiT. The 2.6 GB model.pt is a training checkpoint with EMA and optimizer state)
License: Apache 2.0 weights, MIT code
Languages: Arabic and English, zero-shot cloning
Streaming: No native streaming (CFM decoder)
TTFB / first-chunk: not published
CPU real-time: not published (GPU only: RTF ~0.12, vendor, RTX 4090)
Quantized formats: not published
Integration path: pip install silma-tts; Gradio app. Vocos vocoder pulled from the hub

Open model ↗ Correct this row

Freya-TTS

Turkish, single fixed speaker, no cloning

183M

Apache 2.0

▶

Params: 183,198,145 (safetensors.total)
License: Apache 2.0
Languages: Turkish, single fixed speaker, no cloning
Streaming: No native streaming (32-step Euler ODE over the whole utterance)
TTFB / first-chunk: not published
CPU real-time: not published
Quantized formats: not published
Integration path: DIY; github.com/freyavoiceai/FreyaTTS, install from source. There is no "-small" variant; the org ships one checkpoint

Open model ↗ Correct this row

hindi-tts-190M

Hindi, one fixed reference voice (swappable)

190M

MIT CPU no

▶

Params: ~190,178,332 (derived from the fp32 EMA checkpoint) + bundled Vocos ~13.6M
License: MIT (IndicF5 / F5-TTS derivative; upstream terms also apply)
Languages: Hindi, one fixed reference voice (swappable)
Streaming: No native streaming (flow matching, NFE 32)
TTFB / first-chunk: not published
CPU real-time: no: RTF 3.13 median, 3.13x slower than real time (vendor eval, Apple M4 16 GB, MPS, fp32, NFE 32). No CPU figure published
Quantized formats: not published (fp32 only)
Integration path: Weak: no pip package and no GitHub repo. Inference code must be vendored out of the author's Gradio Space

Open model ↗ Correct this row

Chatterbox-Nano

English (the 23-language list on the card belongs to Chatterbox-Multilingual)

217M

MIT CPU yes

▶

Params: 217,471,139 T3 backbone (reconciled from the fp32 weight file); ~483.0M with s3gen and the voice encoder. The card's "110M" is wrong by ~2x against its own file
License: MIT
Languages: English (the 23-language list on the card belongs to Chatterbox-Multilingual)
Streaming: No native streaming in the official repo
TTFB / first-chunk: not published (the sub-200 ms figure on the card is Resemble's paid hosted service)
CPU real-time: yes: 3x real time on 8 cores (vendor, CPU model and thread policy unspecified)
Quantized formats: not published
Integration path: DIY from GitHub source only: the PyPI package has no nano path. Perth watermarking built in. livekit-plugins-resemble and Pipecat resembleai target the hosted API, not these weights

Open model ↗ Correct this row

NeuTTS Nano

English, plus official German, Spanish and French checkpoints

229M

Revenue-capped

▶

Restriction: Revenue cap: a commercial license is required above $5,000,000 annual revenue
Params: 228,704,832 (safetensors.total)
License: NeuTTS Open License v1.0 (cardData says only "other", so the restriction is invisible from card metadata)
Languages: English, plus official German, Spanish and French checkpoints
Streaming: not published
TTFB / first-chunk: not published
CPU real-time: not published (vendor publishes Q4_0 llama-bench tokens/s only, codec excluded, no RTF)
Quantized formats: GGUF Q4 and Q8, first-party, for every language checkpoint
Integration path: DIY; the neutts package with llama-cpp-python

Open model ↗ Correct this row

NeuTTS-2E

English. 4 fixed speakers, 7 emotions, no cloning

236M

Revenue-capped CPU borderline output stream

▶

Restriction: Revenue cap: a commercial license is required above $5,000,000 annual revenue
Params: 236,039,680 (safetensors.total; ~125M active)
License: NeuTTS Open License v1.0 weights, Apache-2.0 code
Languages: English. 4 fixed speakers, 7 emotions, no cloning
Streaming: Output-streaming, GGUF path only
TTFB / first-chunk: not published
CPU real-time: borderline: every published CPU number is Q4_0, measured on sibling checkpoints, and excludes the codec (vendor, AMD Ryzen 9 HX 370 and iMac M4, llama-bench)
Quantized formats: GGUF Q4 and Q8 (first-party); first-party ONNX codec decoders including int8
Integration path: DIY. livekit-plugins-neuphonic and Pipecat neuphonic wrap the hosted cloud API, not these weights

Open model ↗ Correct this row

Vosk-TTS ru 0.10-multi

Russian, 57 speakers

~ 237M

Apache 2.0 CPU yes

▶

Params: ~236.7M (derived: ~58.7M TTS ONNX + 178M ruBert-base front end)
License: Apache 2.0
Languages: Russian, 57 speakers
Streaming: No native streaming
TTFB / first-chunk: not published
CPU real-time: yes: xRT 0.35 (vendor, hardware unspecified; README inside the model zip)
Quantized formats: ONNX (the only shipped format; no int8 variant)
Integration path: pip install vosk-tts, plus CLI and a server directory. Pure CPU, fully offline. Despite the name it is a DiT diffusion model, not VITS

Open model ↗ Correct this row

ChatTTS

English, Chinese

~ 304M

Non-commercial CPU no output stream

▶

Restriction: Non-commercial weights and AGPL-3.0 code (copyleft, relevant if linked into a service)
Params: ~304.4M (measured across five fp32 shards; no aggregate safetensors index)
License: CC-BY-NC-4.0 weights, AGPL-3.0 code
Languages: English, Chinese
Streaming: Output-streaming
TTFB / first-chunk: not published
CPU real-time: no (>=4 GB VRAM stated; the only published figure is RTF ~0.3 on a 4090)
Quantized formats: not published
Integration path: DIY; Python API and bundled WebUI. The authors deliberately degraded fidelity in training to deter misuse, which is a hard quality ceiling. Weights frozen 2024-10-22

Open model ↗ Correct this row

Vui Nano

English

~ 305M

Unresolved * CPU no sentence stream

▶

Restriction: License unresolved upstream. Do not assert Apache 2.0 on the weights
Params: ~305M (derived at fp16; no safetensors index)
License: GitHub license detection returns NOASSERTION and the two READMEs contradict each other on whether Apache 2.0 covers the weights. The codec and speaker encoder are Alibaba property under the Qwen3-TTS terms
Languages: English
Streaming: Sentence-level (WebRTC and WebSocket with backpressure)
TTFB / first-chunk: not published
CPU real-time: no (GPU required, ~12 GB VRAM for the full assistant stack)
Quantized formats: not published
Integration path: Ships an OpenAI Realtime API compatible server, so it drops into any client speaking that protocol. Weights live under the different HF org slug fluxions

Open model ↗ Correct this row

VoiceCraft giga330M

English

323M

Non-commercial * CPU no

▶

Restriction: Non-commercial, both halves
Params: 323,434,522 (safetensors.total; "330M" is a repo name)
License: Coqui Public Model License 1.0.0 weights, CC BY-NC-SA 4.0 code. The parent repo's cardData contradicts LICENSE-MODEL; the per-checkpoint repos carry no license key at all
Languages: English
Streaming: No native streaming
TTFB / first-chunk: not published
CPU real-time: no (no CPU path or benchmark; CUDA-oriented throughout)
Quantized formats: not published
Integration path: DIY; Gradio app and notebooks. Prompt plus generation capped at 16 s on the TTS-finetuned checkpoint. Weight repos frozen 2024

Open model ↗ Correct this row

Plapre Nano v2

Danish

335M

CC-BY-4.0 * CPU no sentence stream

▶

Params: 335,010,240 (safetensors.total)
License: CC-BY-4.0 (model card only; the GitHub repo has no LICENSE file)
Languages: Danish
Streaming: Sentence-level (raw 24 kHz PCM, chunked transfer)
TTFB / first-chunk: not published
CPU real-time: no (CUDA GPU required; vLLM pinned to >=0.15,<0.16)
Quantized formats: not published
Integration path: DIY; OpenAI-shaped /v1/audio/speech via plapre-serve. The Kanade tokenizer is a separate ~400 MiB GPU-resident dependency

Open model ↗ Correct this row

F5-TTS

English, Mandarin, with code-switching

~ 336M

Non-commercial CPU no

▶

Restriction: Non-commercial weights (inherited from the Emilia training set, so unlikely to be relaxed)
Params: 335.8M (paper; no safetensors index)
License: CC-BY-NC-4.0 weights, MIT code
Languages: English, Mandarin, with code-switching
Streaming: No native streaming (chunk inference is text splitting)
TTFB / first-chunk: not published (the 253 ms figure is full-utterance request latency under Triton + TensorRT-LLM: vendor, one NVIDIA L20)
CPU real-time: no (every first-party benchmark is L20 GPU; 16 to 32 NFE steps plus a separate vocoder)
Quantized formats: No official GGUF or ONNX; in-repo TensorRT-LLM engine recipe
Integration path: Most production-ready path in this table: PyPI package, CLI, Gradio, GHCR image, in-repo Triton runtime. Vocos vocoder is a separate dependency

Open model ↗ Correct this row

Raon-OpenTTS-0.3B

English, 16 kHz output

~ 336M

Non-commercial

▶

Restriction: Non-commercial weights
Params: 336M (model-card prose only; safetensors is null)
License: CC-BY-NC-4.0 weights, Apache-2.0 code
Languages: English, 16 kHz output
Streaming: No native streaming (F5-TTS style flow-matching DiT)
TTFB / first-chunk: not published
CPU real-time: not published
Quantized formats: not published
Integration path: DIY; F5-TTS-compatible code plus a separately fetched HiFi-GAN. The only downloadable artifact is a 5.4 GB raw training checkpoint with no safetensors and no config.json

Open model ↗ Correct this row

IndicF5

11 Indian languages

351M

MIT * gated

▶

Params: 350,681,834 (safetensors.total)
License: MIT (HF card only; the GitHub repo ships no LICENSE file)
Languages: 11 Indian languages
Streaming: No native streaming (flow matching)
TTFB / first-chunk: not published
CPU real-time: not published
Quantized formats: not published
Integration path: DIY; Transformers with trust_remote_code. Needs a reference clip plus its transcript on every call; Vocos vocoder pulled at runtime. HF gated (auto)

Open model ↗ Correct this row

NVIDIA Magpie TTS Multilingual 357M

12: ar, de, en, es, fr, hi, it, ja, ko, pt, vi, zh. 5 fixed English voices, cloning removed in this release

~ 364M

NVIDIA OML

▶

Restriction: NVIDIA Open Model License: commercial use permitted but with attribution, Trustworthy AI compliance and automatic termination. The card also carries a non-commercial-only checkbox that contradicts the license label
Params: 364M (card prose; the repo name says 357m; safetensors is null so no authoritative count exists)
License: NVIDIA Open Model License
Languages: 12: ar, de, en, es, fr, hi, it, ja, ko, pt, vi, zh. 5 fixed English voices, cloning removed in this release
Streaming: No native streaming (standard mode caps at 20 s; long-form is a sliding window)
TTFB / first-chunk: not published
CPU real-time: not published
Quantized formats: f16 GGUF, first-party, but built from an older revision than the .nemo
Integration path: NeMo Speech for the .nemo, NeMo-Speech.cpp for the GGUF. No safetensors and no Transformers-compatible files. livekit-plugins-nvidia and Pipecat nvidia target Riva/NIM. NanoCodec is a separate download

Open model ↗ Correct this row

Marvis-TTS-100m-v0.2

English, French, German

368M

Apache 2.0 CPU borderline output stream

▶

Params: 368,273,408 (safetensors.total; 100M backbone + 60M decoder + bundled Mimi codec)
License: Apache 2.0
Languages: English, French, German
Streaming: Output-streaming (chunked)
TTFB / first-chunk: not published
CPU real-time: borderline (MLX on Apple Silicon GPU; no CPU RTF published)
Quantized formats: MLX 6-bit, 8-bit (first-party sibling repos)
Integration path: DIY; mlx-audio or Transformers

Open model ↗ Correct this row

KaniTTS 400m-en

English; the sibling kani-tts-370m covers en, de, ar, zh, es, ko

370M

Revenue-capped CPU no output stream TTFB

▶

Restriction: Revenue cap: a commercial license is required above $10,000,000 annual revenue
Params: 369,847,040 (safetensors.total; the repo name rounds up)
License: LFM Open License v1.0 weights (cardData says only "other"), Apache-2.0 code, NVIDIA Open Model License on the required NanoCodec
Languages: English; the sibling kani-tts-370m covers en, de, ar, zh, es, ko
Streaming: Output-streaming (SSE) via the separate kanitts-vllm server
TTFB / first-chunk: first chunk <300 ms in streaming mode (vendor, RTX 5090, vLLM; stated as an upper bound, not a measured percentile)
CPU real-time: no (all published figures are GPU; the slowest is RTF 0.600 on an RTX 3060)
Quantized formats: MLX and MLX 8-bit (vendor); community GGUF
Integration path: kanitts-vllm exposes OpenAI-compatible /v1/audio/speech, drivable from Pipecat or LiveKit OpenAI TTS clients

Open model ↗ Correct this row

KaniTTS-2-en

English

370M

Revenue-capped * CPU no output stream

▶

Restriction: Revenue cap: a commercial license is required above $10,000,000 annual revenue
Params: 369,977,088 (safetensors.total; the card claims 400M)
License: LFM Open License v1.0 weights; the kani-tts-2 code repo has no LICENSE file; NVIDIA OML on the NanoCodec
Languages: English
Streaming: Output-streaming via the separate OpenAI-compatible server
TTFB / first-chunk: not published
CPU real-time: no: ~20 s to 60 s for 10 s of audio (vendor, self-described as not production-ready)
Quantized formats: Community only: GGUF, MLX, bitsandbytes 4-bit
Integration path: DIY; PyPI package plus the vendor server. Requires pinned nemo-toolkit and transformers versions

Open model ↗ Correct this row

VietTTS

Vietnamese

440M

Non-commercial

▶

Restriction: Non-commercial weights
Params: 440,482,096 core (measured: llm 315,146,753 + flow 104,874,752 + hift 20,460,591)
License: CC BY-NC weights (cardData is the bare string "cc"), Apache-2.0 code
Languages: Vietnamese
Streaming: No true incremental synthesis (the streaming response is chunked HTTP delivery of a finished WAV)
TTFB / first-chunk: not published
CPU real-time: not published
Quantized formats: not published
Integration path: Best integration path in this table: OpenAI-compatible /v1/audio/speech, /v1/voices, Docker and docker-compose, plus a CLI. Last substantive commit 2024-12-12

Open model ↗ Correct this row

MegaTTS 3

Chinese, English, code-switching

~ 450M

Apache 2.0 CPU no

▶

Params: 0.45B diffusion transformer backbone (the full download is ~1.31B across 5 components)
License: Apache 2.0
Languages: Chinese, English, code-switching
Streaming: No native streaming
TTFB / first-chunk: not published
CPU real-time: no: ~30 s for 10 inference steps (vendor, hardware not named)
Quantized formats: not published
Integration path: DIY; Gradio demo and inference script. The WaveVAE encoder is deliberately withheld, so cloning a new speaker requires submitting a clip to ByteDance and waiting for a published latent

Open model ↗ Correct this row

VoXtream2

English

463M

CC-BY-4.0 * CPU no dual stream TTFB

▶

Params: 462,870,528 (safetensors.total); + Kyutai Mimi codec 96,151,393 at runtime
License: CC-BY-4.0 weights. Code license is contradictory: GitHub API says Apache-2.0, the README says MIT
Languages: English
Streaming: Dual-streaming (word-by-word text in, 80 ms audio chunks out)
TTFB / first-chunk: v1: 102 ms first packet (vendor, A100, compiled, FP16) and 123 ms (vendor, RTX 3090, compiled). v2 claims 74 ms with no hardware named
CPU real-time: no: ~2.2 GB VRAM, no CPU mode documented
Quantized formats: not published
Integration path: pip install voxtream; CLI, Python generator, WebSocket server, Gradio. Needs espeak-ng. Reuses the Sesame CSM depth transformer (architectural reuse, not a fine-tune)

Open model ↗ Correct this row

GPT-SoVITS v4

Chinese, English, Japanese, Korean, Cantonese

~ 491M

MIT output stream

▶

Params: ~491M model (derived: SoVITS ~384.5M + GPT ~77.6M + vocoder ~28.9M; no published count). The mandatory roberta and HuBERT front ends add ~420M more
License: MIT (code and weights)
Languages: Chinese, English, Japanese, Korean, Cantonese
Streaming: Output-streaming (chunked, api_v2.py, 4 quality/latency levels)
TTFB / first-chunk: not published
CPU real-time: not published for v4 (the RTF 0.526 figure is published for v2 ProPlus only: vendor, Apple M4 CPU)
Quantized formats: not published
Integration path: DIY; in-repo FastAPI server with a chunked streaming endpoint, plus Gradio WebUI. Upstream now positions v2Pro as surpassing v4 at lower cost

Open model ↗ Correct this row

Chatterbox

23 languages (Multilingual v3)

~ 500M

MIT CPU borderline output stream TTFB

▶

Params: ~0.5B
License: MIT (code and weights)
Languages: 23 languages (Multilingual v3)
Streaming: Output-streaming
TTFB / first-chunk: low latency claimed for Turbo variant (vendor; hardware unspecified)
CPU real-time: borderline
Quantized formats: not published
Integration path: Community wrappers; built-in PerTh watermarking

Open model ↗ Correct this row

CosyVoice 2

Zh, En, + others; cross-lingual mode

~ 500M

Apache 2.0 CPU no dual stream TTFB

▶

Params: 0.5B
License: Apache 2.0
Languages: Zh, En, + others; cross-lingual mode
Streaming: Dual-capable chunk-aware streaming
TTFB / first-chunk: ~150 ms first-chunk (vendor paper; GPU)
CPU real-time: no
Quantized formats: not published
Integration path: Community wrappers. The old FunAudioLLM GitHub org was renamed; that URL now redirects

Open model ↗ Correct this row

spark-tts-salt

7 East African: Acholi, Ateso, Ugandan English, Luganda, Lugbara, Runyankore, Swahili

507M

Non-commercial * output stream gated

▶

Restriction: The declared MIT grant is invalid: derived from CC-BY-NC-SA-4.0 base weights, and the required BiCodec is also NC-SA. ShareAlike cannot be relicensed downstream
Params: 506,634,112 (safetensors.total), LLM only
License: Declares mit. Both the declared base and the original upstream are cc-by-nc-sa-4.0. Effective runtime license is NC-SA
Languages: 7 East African: Acholi, Ateso, Ugandan English, Luganda, Lugbara, Runyankore, Swahili
Streaming: Output-streaming only via the upstream Triton runtime; this repo ships no streaming code
TTFB / first-chunk: not published
CPU real-time: not published
Quantized formats: not published
Integration path: Requires cloning the upstream Spark-TTS for the BiCodec encode and decode path: this repo ships the LLM only and cannot synthesize standalone. HF gated (auto)

Open model ↗ Correct this row

Gepard-1.0

en-US, en-GB, es-MX, pt-BR, nl-NL

556M

Apache 2.0 * CPU no output stream TTFB

▶

Params: 555,694,169 (safetensors.total)
License: Apache 2.0, but the pipeline requires the NVIDIA NeMo NanoCodec under the NVIDIA Open Model License
Languages: en-US, en-GB, es-MX, pt-BR, nl-NL
Streaming: Output-streaming (chunked, vLLM)
TTFB / first-chunk: ~50 ms TTFA (vendor, single RTX 5090, served on vLLM; ~25x real time)
CPU real-time: no (CUDA + vLLM required)
Quantized formats: not published
Integration path: DIY; gepard-inference. The hosted API is Cartesia-wire-compatible; the self-hosted weights are not

Open model ↗ Correct this row

Marvis-TTS-250m-v0.2

English, French, German

569M

Apache 2.0 CPU borderline output stream

▶

Params: 568,866,816 (safetensors.total; 250M backbone + 60M decoder + bundled Mimi codec)
License: Apache 2.0
Languages: English, French, German
Streaming: Output-streaming (chunked)
TTFB / first-chunk: not published
CPU real-time: borderline (MLX on Apple Silicon GPU; no CPU RTF published)
Quantized formats: MLX 4-bit, 6-bit, 8-bit (first-party sibling repos)
Integration path: DIY; mlx-audio or Transformers CsmForConditionalGeneration. Cloning via a reference clip

Open model ↗ Correct this row

Audio8-TTS-Preview-0.6b

11: yue, zh, nl, en, fr, de, it, ja, ko, pl, es

601M

Apache 2.0 output stream

▶

Params: 601,159,424 backbone (safetensors.total); bundled codec ~337.4M is unverified (pickle with no index)
License: Apache 2.0
Languages: 11: yue, zh, nl, en, fr, de, it, ja, ko, pl, es
Streaming: Output-streaming (SSE, ~0.56 s per chunk)
TTFB / first-chunk: not published (the published figure is warm p50 end-to-end 0.691 s, RTF 0.116: vendor, one NVIDIA H20)
CPU real-time: not published (an INT4 ONNX Runtime CPU path is documented with no RTF attached)
Quantized formats: ONNX INT4 (vendor). The MLX build is community, not first-party
Integration path: The SGLang Omni adapter serves OpenAI-shaped /v1/audio/speech with SSE, so it drives from LiveKit or Pipecat OpenAI TTS clients

Open model ↗ Correct this row

OuteTTS-1.0-0.6B

602M

Apache 2.0

▶

Params: 602,013,696 (safetensors.total)
License: Apache 2.0. The IBM Research DAC.speech.v1.0 codec has its own license
Languages: 14
Streaming: No native streaming (chunked and batch modes split text, they do not emit audio incrementally)
TTFB / first-chunk: not published
CPU real-time: not published (a llama.cpp CPU path is documented, no CPU RTF from vendor or community)
Quantized formats: GGUF, ONNX, FP8, EXL2 (all first-party); community MLX
Integration path: DIY via the outetts package (Transformers, llama.cpp, ExLlamaV2, vLLM backends). Do not substitute Llama-OuteTTS-1.0-1B: 1,248,397,312 params and CC-BY-NC-SA-4.0

Open model ↗ Correct this row

Indic-Mio

22 scheduled Indian languages + English

609M

Apache 2.0 CPU no

▶

Params: 608,894,976 (safetensors.total)
License: Apache 2.0. The required MioCodec is Aratako/MioCodec-25Hz-24kHz, third-party, MIT
Languages: 22 scheduled Indian languages + English
Streaming: not published
TTFB / first-chunk: not published
CPU real-time: no (both documented paths are vLLM or Transformers with device_map cuda)
Quantized formats: not published (community MLX fp16)
Integration path: DIY; vllm serve plus the community MioTTS-Inference server. Derivative of MioTTS-0.6B

Open model ↗ Correct this row

MioTTS-0.6B

English, Japanese

609M

Apache 2.0

▶

Params: 608,894,976 (safetensors.total)
License: Apache 2.0 (the GGUF sibling repo declares "other"). Bundled reference-voice presets are not licensed for commercial use: supply your own reference audio
Languages: English, Japanese
Streaming: No native audio streaming (LM token streaming works via OpenAI-compatible servers; codec decode is not incremental)
TTFB / first-chunk: not published
CPU real-time: not published
Quantized formats: GGUF BF16, Q8_0, Q6_K, Q4_K_M (first-party, Aratako/MioTTS-GGUF)
Integration path: Low friction: the LM is an ordinary causal LM, so it serves from vLLM, llama.cpp, Ollama or SGLang, plus the MIT MioTTS-Inference wrapper

Open model ↗ Correct this row

OmniVoice

600+

613M

Non-commercial

▶

Restriction: Non-commercial weights (inherited from Emilia), stated only in the README. High-traffic community GGUF and MLX mirrors are mistagged apache-2.0, so downstream users believe they hold a commercial license they do not
Params: 612,577,288 (safetensors.total) + a required 201,400,553 audio tokenizer
License: CC-BY-NC weights per README prose; cardData.license is null. Code Apache-2.0
Languages: 600+
Streaming: No native streaming
TTFB / first-chunk: not published
CPU real-time: not published (the README claims RTF as low as 0.025 with no hardware named, so it is not publishable as a number)
Quantized formats: Community: MLX 4-bit, 8-bit, bfloat16; LiteRT conversions. An official ONNX export is an open feature request
Integration path: pip install omnivoice. Deep community ecosystem: vLLM-Omni, RealtimeTTS, audio.cpp, mlx-audio, an OpenAI-compatible shim. The highest-traffic sub-1B open TTS model found in this sweep

Open model ↗ Correct this row

Malaysian-TTS-0.6B-v1

Malay, English, Singlish; 7 voices

629M

No grant *

▶

Restriction: No license stated anywhere. No rights granted, and the required DistilCodec detokenizer is separately CC-BY-NC-4.0, so the pipeline is non-commercial regardless
Params: 629,342,208 (safetensors.total), LM only
License: none (cardData.license is null, no license tag, no LICENSE file); training and inference code MIT; required codec CC-BY-NC-4.0
Languages: Malay, English, Singlish; 7 voices
Streaming: No native streaming
TTFB / first-chunk: not published
CPU real-time: not published
Quantized formats: not published
Integration path: DIY; Transformers plus a manual DistilCodec install and a manual fetch of two codec files. Not runnable standalone. Requires upstream text normalization. Its pipeline_tag is text-generation, so it does not appear in HF TTS listings

Open model ↗ Correct this row

VoxCPM-0.5B

Chinese, English

~ 652M

Apache 2.0 CPU no output stream

▶

Params: ~652M core (derived at bfloat16) + a 301 MB AudioVAE. "0.5B" is the MiniCPM-4-0.5B backbone name, not a measurement
License: Apache 2.0 (weights and code, explicitly)
Languages: Chinese, English
Streaming: Output-streaming (generate_streaming)
TTFB / first-chunk: not published
CPU real-time: no (CUDA >= 12.0; the device is hardcoded to cuda in config)
Quantized formats: not published (community GGUF, ONNX, RKNN2, CoreML)
Integration path: DIY; the voxcpm package and CLI. Superseded upstream: the repo now tracks VoxCPM2 (2B, over the line) and this checkpoint has not moved since 2025-09-19

Open model ↗ Correct this row

NeuTTS Air

English

748M

Apache 2.0 * CPU yes output stream

▶

Params: 747,930,496 (safetensors.total)
License: Apache 2.0. Note the newer neutts-2e and neutts-nano siblings moved to a revenue-capped license; Air did not
Languages: English
Streaming: Output-streaming
TTFB / first-chunk: not published
CPU real-time: yes (GGUF on-device is the design target)
Quantized formats: GGUF
Integration path: DIY; built for edge

Open model ↗ Correct this row

Irodori-TTS-v4-Small

Japanese

766M

MIT

▶

Params: 766,052,385 (safetensors.total; the 3.06 GB download is fp32 storage, not a 3B model. The HF widget rounds to 0.8B)
License: MIT (weights and code)
Languages: Japanese
Streaming: No native streaming (duration predictor then whole-utterance rectified-flow DiT)
TTFB / first-chunk: not published
CPU real-time: not published (cpu-only and macOS MPS install extras exist; no RTF on any hardware)
Quantized formats: torchao INT8, INT4, FP8 (vendor, separate repo)
Integration path: DIY; PyPI package. Requires the separate Semantic-DACVAE-Japanese-32dim codec. SilentCipher watermark applied when the dependency is present

Open model ↗ Correct this row

Kyutai TTS 0.75B

English

~ 780M

CC-BY-4.0 CPU no dual stream

▶

Params: ~780M (derived at bf16; the "750M" is name-derived, no safetensors index). Mimi tokenizer is a separate ~192M checkpoint
License: CC-BY-4.0 weights. Code is dual: MIT for the Python parts, Apache-2.0 for the Rust backend
Languages: English
Streaming: Dual: text consumed incrementally, audio shifted 16 steps (1.28 s) at 12.5 Hz
TTFB / first-chunk: not published for this checkpoint (the 220 ms figure belongs to Kyutai's 2B-class announcement model)
CPU real-time: no (PyTorch/CUDA, Rust CUDA server, or MLX on Apple Silicon)
Quantized formats: No artifact; MLX runtime quantize flags only
Integration path: DIY; the moshi package, a Rust websocket server, MLX scripts. Community LiveKit plugin only, no Pipecat service

Open model ↗ Correct this row

Sarashina2.2-TTS

Japanese primary, English secondary, code-switching

810M

Non-commercial CPU no

▶

Restriction: Non-commercial, plus a mandatory inaudible watermark on every output, a name prefix on derivatives, an attribution notice, and a licensor right to revise the terms after download
Params: 809,908,480 LM (safetensors.total); flow, hift and campplus add ~140M more
License: Sarashina Model NonCommercial License Agreement v2.0 (LICENSE file text; cardData.license is null, so HF facet search will not surface it)
Languages: Japanese primary, English secondary, code-switching
Streaming: No native streaming
TTFB / first-chunk: not published
CPU real-time: no (GPU required; ~6 GB VRAM on the Transformers backend)
Quantized formats: not published
Integration path: Well packaged: pip-installable, Gradio UI, Dockerfile, optional vLLM backend. HF only, no GitHub repo

Open model ↗ Correct this row

VoiceCraft giga830M

English

840M

Non-commercial CPU no

▶

Restriction: Non-commercial, both halves
Params: 839,565,338 (safetensors.total; "830M" is a repo name)
License: Coqui Public Model License 1.0.0 weights, CC BY-NC-SA 4.0 code. This repo carries no license key in cardData
Languages: English
Streaming: No native streaming
TTFB / first-chunk: not published
CPU real-time: no
Quantized formats: not published
Integration path: DIY; Gradio app and notebooks. Frozen 2024-04-16

Open model ↗ Correct this row

Fun-CosyVoice3-0.5B-2512

9 + 18 Chinese dialects

~ 859M

Apache 2.0 dual stream TTFB

▶

Params: ~859M (derived: llm ~506M + flow ~332M + hift ~21M at fp32; "0.5B" describes the LLM only)
License: Apache 2.0 (weights and code)
Languages: 9 + 18 Chinese dialects
Streaming: Dual-capable: text-in streaming and audio-out streaming
TTFB / first-chunk: ~150 ms (vendor card, hardware unspecified)
CPU real-time: not published
Quantized formats: not published (the in-repo ONNX files are fp32 pipeline components)
Integration path: DIY via QwenAudio/CosyVoice, vLLM supported. Pipecat's funasr service is FunAudioLLM ASR, not this

Open model ↗ Correct this row

OpenAudio S1-mini

868M

Non-commercial gated

▶

Restriction: Non-commercial
Params: 867,561,487 (derived from model.pth at bf16; the vendor's "0.5B" is off by ~74 percent). The required codec adds an undetermined amount
License: cc-by-nc-sa-4.0 on the HF card; the fish-speech README instead names the "Fish Audio Research License". Both non-commercial
Languages: 13
Streaming: not published for S1-mini (the SGLang streaming work is scoped to S2-Pro)
TTFB / first-chunk: not published (the ~100 ms TTFA figure is S2-Pro 4B on an H200)
CPU real-time: not published
Quantized formats: not published
Integration path: DIY via fish-speech. HF gated (auto) with a non-commercial acknowledgment checkbox. LiveKit's Fish Audio plugin and Pipecat fish target the hosted cloud API

Open model ↗ Correct this row

Parler-TTS Mini

English

~ 880M

Apache 2.0 CPU no output stream

▶

Params: 880M
License: Apache 2.0
Languages: English
Streaming: Output-streaming supported
TTFB / first-chunk: not published
CPU real-time: no
Quantized formats: not published
Integration path: DIY; HF Transformers native. Rule 5 flag: checkpoint frozen 2024, repo last pushed 2024-12-10 with no maintainer commits since. Archive candidate

Open model ↗ Correct this row

Qwen3-TTS-12Hz-0.6B-CustomVoice

10; 9 fixed premium timbres with instruction control, no reference-audio cloning (that is the Base checkpoint)

906M

Apache 2.0 CPU no output stream

▶

Params: 905,788,672 LM (safetensors.total); the required speech tokenizer adds ~170.5M
License: Apache 2.0
Languages: 10; 9 fixed premium timbres with instruction control, no reference-audio cloning (that is the Base checkpoint)
Streaming: Output-streaming (chunked)
TTFB / first-chunk: not published (the 97 ms claim names no hardware, batch size or conditions)
CPU real-time: no
Quantized formats: not published first-party (community MLX, GGUF, ONNX, OpenVINO INT8)
Integration path: DIY; QwenLM/Qwen3-TTS, vLLM-Omni offline inference. Genuine open weights, not the hosted Qwen3-TTS-Flash API

Open model ↗ Correct this row

Qwen3-TTS-12Hz-0.6B-Base

10 + Chinese dialect voices

915M

Apache 2.0 CPU no output stream

▶

Params: 914,643,008 LM (safetensors.total) + speech tokenizer 170,557,441
License: Apache 2.0
Languages: 10 + Chinese dialect voices
Streaming: Output-streaming (chunked)
TTFB / first-chunk: not published (same 97 ms claim, no hardware)
CPU real-time: no
Quantized formats: not published first-party (community MLX, GGUF, ONNX, OpenVINO INT8)
Integration path: DIY; QwenLM/Qwen3-TTS, vLLM-Omni offline inference

Open model ↗ Correct this row

Sesame CSM-1B

English

~ 1.00B

Apache 2.0 CPU no other stream

▶

Params: 1B (boundary case, see rule 1)
License: Apache 2.0
Languages: English
Streaming: Conversational, context-conditioned generation
TTFB / first-chunk: not published
CPU real-time: no
Quantized formats: community quantized variants exist; verify
Integration path: Community wrappers

Open model ↗ Correct this row

Piper

Many languages, per-language voices

—

GPL CPU yes sentence stream

▶

Params: tens of M per voice
License: GPL v3
Languages: Many languages, per-language voices
Streaming: Sentence-level streaming
TTFB / first-chunk: not published
CPU real-time: yes
Quantized formats: ONNX (ships as ONNX)
Integration path: DIY; the default for Home Assistant and Raspberry Pi. GPL v3: fine to run as a service, read the license for distribution

Open model ↗ Correct this row

Silero V5 CIS base

20+ CIS and Russia languages, 80+ speakers

—

MIT CPU yes

▶

Params: not published (87.4 MB TorchScript checkpoint bounds it near 22M)
License: MIT (LICENSE_CIS; covers v5_cis_base and v5_cis_base_nostress only, not v5_ru or v5_cis_ext)
Languages: 20+ CIS and Russia languages, 80+ speakers
Streaming: No native streaming (utterance-level apply_tts)
TTFB / first-chunk: not published
CPU real-time: yes: 36.48x real time at 1 thread, 85.72x at 4 threads (vendor wiki, Intel Core i9-10940X @ 3.30 GHz)
Quantized formats: not published
Integration path: torch.hub or pip install silero. Pure PyTorch, vocoder included, no external dependency

Open model ↗ Correct this row

StyleTTS 2

English (single and multi-speaker)

—

MIT CPU borderline

▶

Params: under 1B
License: MIT
Languages: English (single and multi-speaker)
Streaming: No native streaming
TTFB / first-chunk: not published
CPU real-time: borderline
Quantized formats: community ONNX efforts
Integration path: DIY

Open model ↗ Correct this row

MeloTTS

En (4 accents), Es, Fr, Zh, Ja, Ko

—

MIT CPU yes

▶

Params: under 1B (VITS-based)
License: MIT
Languages: En (4 accents), Es, Fr, Zh, Ja, Ko
Streaming: No native streaming; sentence-chunk workaround
TTFB / first-chunk: not published
CPU real-time: yes
Quantized formats: not published
Integration path: DIY; simple Python API

Open model ↗ Correct this row

Silero V5 RU and CIS ext

Russian (v5_ru through v5_5_ru) plus the CIS ext speaker set

—

Non-commercial

▶

Restriction: Non-commercial share-alike. The MIT LICENSE_CIS covers only the v5_cis_base variants, which are in the main table
Params: not published (TorchScript checkpoints, no exact count anywhere)
License: CC-BY-NC-SA-4.0 (main repo LICENSE; GitHub reports NOASSERTION)
Languages: Russian (v5_ru through v5_5_ru) plus the CIS ext speaker set
Streaming: No native streaming
TTFB / first-chunk: not published
CPU real-time: not published (see the CIS base row for the published i9-10940X figures)
Quantized formats: not published
Integration path: torch.hub or pip install silero. Self-contained TorchScript, vocoder included. Note livekit-plugins-silero and Pipecat's Silero service are VAD, not TTS

Open model ↗ Correct this row