{
    "meta": {
        "title": "D-Central — Local Voice-AI Model Database (STT + TTS)",
        "description": "21 self-hostable speech-to-text and text-to-speech models scored on license commercial-safety, size, languages (EN/FR flagged), quality, real-time factor by hardware, streaming, Raspberry-Pi viability and offline operation — the speech layer of a sovereign local-AI stack.",
        "generated": "2026-07-18T07:11:49+00:00",
        "version": "1.0",
        "license": "https://creativecommons.org/licenses/by/4.0/",
        "license_name": "CC BY 4.0",
        "source": "https://d-central.tech/local-voice-ai-models/",
        "method": "The load-bearing column is commercial-safety of the LICENSE, not raw quality: permissive (MIT/Apache/BSD) and attribution (CC-BY-4.0) are commercial-OK; copyleft (GPL-3.0, e.g. the maintained Piper fork) is fine self-hosted but constrains embedding in distributed proprietary software; non-commercial (XTTS-v2 CPML, Fish CC-BY-NC-SA, original Canary CC-BY-NC) is a dead end for a product. Real-time factors are marked unverified where only anecdotal. Verified July 2026 against model repos / Hugging Face cards.",
        "provenance": "Model GitHub repos and Hugging Face model cards; licenses cross-checked against LICENSE files where available. Compiled via a D-Central research pass.",
        "disclaimer": "Licenses and variants change — verify the exact checkpoint before shipping (Canary and Voxtral in particular have per-variant license differences). RTF figures depend heavily on hardware, quantization and audio length. Not legal advice on license compliance."
    },
    "rows": [
        {
            "model": "whisper.cpp",
            "type": "STT",
            "developer": "Georgi Gerganov (ggml-org)",
            "license": "MIT",
            "commercial": "permissive",
            "size": "runtime; GGML models ~75 MB (tiny q) → ~3.1 GB (large-v3)",
            "languages": "multilingual (~99, incl. FR)",
            "quality": "Same accuracy as the Whisper checkpoint it runs (minor loss at q4/q5)",
            "speed": "Faster than realtime on ordinary laptop CPUs (base/small); tiny/base run on Raspberry Pi",
            "streaming": "partial",
            "offline": "yes",
            "min_hw": "CPU-only, down to Pi / phones",
            "pi": true,
            "best_for": "Shipping Whisper inside a product: one C/C++ binary, no Python",
            "grade": "A",
            "notes": "The obvious DCENT_Voice engine. `stream` example is sliding-window pseudo-streaming, not true streaming."
        },
        {
            "model": "faster-whisper",
            "type": "STT",
            "developer": "SYSTRAN (CTranslate2)",
            "license": "MIT",
            "commercial": "permissive",
            "size": "runtime; Whisper weights in CTranslate2 format",
            "languages": "multilingual (incl. FR)",
            "quality": "Same WER as reference Whisper for the same checkpoint",
            "speed": "~4× faster than openai/whisper, less memory; faster-than-realtime large-v3 on mid-range NVIDIA",
            "streaming": "no",
            "offline": "yes",
            "min_hw": "CPU workable (small/int8); 4–8 GB VRAM for large-v3",
            "pi": false,
            "best_for": "Fastest full-accuracy Whisper on a GPU or beefy CPU",
            "grade": "A",
            "notes": "The default server-side runtime; powers most self-hosted transcription stacks."
        },
        {
            "model": "Whisper (tiny→large-v3)",
            "type": "STT",
            "developer": "OpenAI",
            "license": "MIT",
            "commercial": "permissive",
            "size": "39 M (tiny) → 1.55 B (large-v3); turbo 809 M",
            "languages": "multilingual (~99, incl. FR)",
            "quality": "large-v3 ~7.4% avg WER; the multilingual accuracy reference",
            "speed": "Original PyTorch slow; large-v3-turbo ~8× faster than large-v3",
            "streaming": "no",
            "offline": "yes",
            "min_hw": "CPU for tiny/base; ~10 GB VRAM for large-v3 fp16",
            "pi": true,
            "best_for": "Multilingual accuracy baseline; the weights everything else runs or distils",
            "grade": "A",
            "notes": "Weights + code MIT — fully commercial-safe. large-v3-turbo (809 M) is the practical default size."
        },
        {
            "model": "NVIDIA Parakeet TDT 0.6B (v2/v3)",
            "type": "STT",
            "developer": "NVIDIA (NeMo)",
            "license": "CC-BY-4.0",
            "commercial": "attribution",
            "size": "600 M",
            "languages": "v2 English; v3 25 European (incl. FR) + auto-detect",
            "quality": "v2 topped Open ASR Leaderboard at 6.05% avg WER — beats Whisper large-v3 on English",
            "speed": "RTFx ~3380 batched on datacenter GPU; community CPU/GGUF ports faster-than-realtime (unverified)",
            "streaming": "partial",
            "offline": "yes",
            "min_hw": "~2–4 GB VRAM; CPU via community ports",
            "pi": false,
            "best_for": "Best English accuracy-per-watt for batch transcription",
            "grade": "A",
            "notes": "CC-BY-4.0 = commercial OK with attribution. Native runtime is NeMo (heavier than whisper.cpp)."
        },
        {
            "model": "NVIDIA Canary (1b-flash / 1b-v2)",
            "type": "STT",
            "developer": "NVIDIA (NeMo)",
            "license": "CC-BY-4.0 (flash/v2); original canary-1b CC-BY-NC",
            "commercial": "attribution",
            "size": "883 M (flash) / ~1 B (v2)",
            "languages": "flash EN/DE/FR/ES + translation; v2 25 European",
            "quality": "SOTA multitask ASR + speech-translation at release",
            "speed": ">1000 RTFx batched on GPU",
            "streaming": "no",
            "offline": "yes",
            "min_hw": "~4 GB VRAM",
            "pi": false,
            "best_for": "Combined transcription + speech translation (EN↔FR) in one local model",
            "grade": "A",
            "notes": "🔴 License VARIES BY VARIANT: original canary-1b = CC-BY-NC (non-commercial); flash + v2 = CC-BY-4.0. Check the exact checkpoint."
        },
        {
            "model": "Distil-Whisper (large-v3)",
            "type": "STT",
            "developer": "Hugging Face",
            "license": "MIT",
            "commercial": "permissive",
            "size": "756 M",
            "languages": "English-only",
            "quality": "Within ~1% WER of large-v3 on out-of-distribution English",
            "speed": "~6× faster than large-v3; drop-in with faster-whisper/whisper.cpp",
            "streaming": "no",
            "offline": "yes",
            "min_hw": "CPU workable quantized; ~4 GB VRAM",
            "pi": false,
            "best_for": "English-only pipelines wanting large-v3 accuracy at small-model speed",
            "grade": "A",
            "notes": "English only — a real limit for a bilingual EN/FR audience."
        },
        {
            "model": "Kyutai STT (1b-en_fr / 2.6b-en)",
            "type": "STT",
            "developer": "Kyutai Labs",
            "license": "CC-BY-4.0 (weights); code Apache/MIT",
            "commercial": "attribution",
            "size": "1 B (EN+FR) / 2.6 B (EN)",
            "languages": "English + French (1B); English (2.6B)",
            "quality": "Competitive WER with word-level timestamps; built for streaming",
            "speed": "Natively real-time streaming, 0.5 s delay; an H100 serves 400 streams",
            "streaming": "yes",
            "offline": "yes",
            "min_hw": "GPU recommended; 1B on consumer GPUs / Apple Silicon (MLX)",
            "pi": false,
            "best_for": "True low-latency streaming dictation — and EN+FR is exactly the D-Central language pair",
            "grade": "A",
            "notes": "Delayed Streams Modeling (from Moshi). CC-BY-4.0 = commercial OK with attribution."
        },
        {
            "model": "Voxtral Mini 3B",
            "type": "STT",
            "developer": "Mistral AI",
            "license": "Apache-2.0",
            "commercial": "permissive",
            "size": "~3 B (audio-LLM; ~9.5 GB GPU RAM bf16)",
            "languages": "multilingual (strong EN/FR/ES/DE)",
            "quality": "Mistral claims it beats Whisper large-v3 on transcription; also does audio Q&A/summarization",
            "speed": "GPU-class inference; not an edge model",
            "streaming": "no",
            "offline": "yes",
            "min_hw": "~10 GB VRAM (bf16); quantized lower",
            "pi": false,
            "best_for": "Speech UNDERSTANDING (transcribe + summarize + answer) fully locally",
            "grade": "A",
            "notes": "July 2025. Do NOT conflate with the separate 2026 'Voxtral TTS' models reported CC-BY-NC (unverified)."
        },
        {
            "model": "Moonshine (tiny/base)",
            "type": "STT",
            "developer": "Useful Sensors / Moonshine AI",
            "license": "MIT",
            "commercial": "permissive",
            "size": "27 M (tiny) / 61 M (base)",
            "languages": "English-only",
            "quality": "Beats Whisper tiny/base WER at comparable size; compute scales with clip length",
            "speed": "~5× faster than Whisper equivalents on short clips; runs on Pi-class edge",
            "streaming": "yes",
            "offline": "yes",
            "min_hw": "CPU-only, Raspberry Pi / edge",
            "pi": true,
            "best_for": "Sub-second-latency voice commands and live captions on tiny hardware",
            "grade": "A",
            "notes": "Moonshine v2 (2026) adds a streaming encoder. English-only is the tradeoff."
        },
        {
            "model": "Vosk",
            "type": "STT",
            "developer": "Alpha Cephei (Kaldi)",
            "license": "Apache-2.0",
            "commercial": "permissive",
            "size": "small ~50 MB, large ~1–2 GB per language",
            "languages": "multilingual (20+, incl. FR)",
            "quality": "Worse WER than Whisper-class, but respectable for its size",
            "speed": "Real-time streaming on CPU incl. Pi and Android",
            "streaming": "yes",
            "offline": "yes",
            "min_hw": "CPU-only, Pi / phones",
            "pi": true,
            "best_for": "True streaming with partial results on the weakest hardware; mature bindings",
            "grade": "A",
            "notes": "Pre-transformer; pick it for latency and footprint, not accuracy."
        },
        {
            "model": "WhisperX",
            "type": "STT",
            "developer": "Max Bain (m-bain)",
            "license": "BSD-2-Clause",
            "commercial": "permissive",
            "size": "pipeline on faster-whisper + alignment + diarization",
            "languages": "multilingual (alignment per-language)",
            "quality": "Whisper accuracy + accurate word-level timestamps via forced alignment",
            "speed": "~70× realtime batched large-v2 on GPU",
            "streaming": "no",
            "offline": "yes",
            "min_hw": "~8 GB VRAM for full diarization",
            "pi": false,
            "best_for": "Word-level timestamps and speaker diarization (meetings, subtitles)",
            "grade": "A",
            "notes": "Diarization uses pyannote models needing HF gated-terms acceptance (still free/offline after download)."
        },
        {
            "model": "Kokoro-82M",
            "type": "TTS",
            "developer": "hexgrad",
            "license": "Apache-2.0",
            "commercial": "permissive",
            "size": "82 M (~330 MB fp32)",
            "languages": "primarily English (strong); 8 langs thinner (FR limited)",
            "quality": "Punches far above its weight — topped community TTS rankings vs far larger models",
            "speed": "Faster than realtime on modern desktop CPUs; trivially realtime on any GPU",
            "streaming": "partial",
            "offline": "yes",
            "min_hw": "CPU-only (desktop); ~1–3 GB RAM",
            "pi": false,
            "best_for": "Best permissive-licensed quality-per-FLOP — the default for commercial local narration",
            "grade": "A",
            "notes": "Apache-2.0 weights trained only on permissive audio. Fixed voice pack, no cloning. Pi 5 borderline (unverified)."
        },
        {
            "model": "Chatterbox / Multilingual",
            "type": "TTS",
            "developer": "Resemble AI",
            "license": "MIT",
            "commercial": "permissive",
            "size": "0.5 B",
            "languages": "English (base); Multilingual 23+ incl. FR",
            "quality": "Beat ElevenLabs in blind preference (~63%, vendor eval); emotion-exaggeration control",
            "speed": "Realtime on mid-range NVIDIA (A10/3060); CPU slow (unverified)",
            "streaming": "partial",
            "offline": "yes",
            "min_hw": "~6 GB VRAM; CPU/ROCm via community",
            "pi": false,
            "best_for": "The best MIT-licensed zero-shot voice cloning — the commercial-safe XTTS replacement",
            "grade": "A",
            "notes": "Output carries Resemble's PerTh neural watermark baked in — transparency plus or caveat depending on use."
        },
        {
            "model": "Piper",
            "type": "TTS",
            "developer": "Rhasspy → Open Home Foundation",
            "license": "MIT (archived) / GPL-3.0 (maintained piper1-gpl)",
            "commercial": "copyleft",
            "size": "~15–60 MB per voice",
            "languages": "multilingual (30+, incl. FR)",
            "quality": "Clearly synthetic but pleasant and highly intelligible; the embedded/assistant standard",
            "speed": "Faster than realtime on Raspberry Pi 4; near-instant on desktop CPU",
            "streaming": "partial",
            "offline": "yes",
            "min_hw": "CPU-only, Raspberry Pi",
            "pi": true,
            "best_for": "Instant offline speech on the weakest hardware; voice-assistant responses",
            "grade": "A",
            "notes": "🔴 License split: original MIT repo archived Oct 2025; active dev (piper1-gpl) is GPL-3.0 (embeds espeak-ng). Fine self-hosted; matters if embedding in distributed proprietary software. Per-voice dataset licenses vary."
        },
        {
            "model": "MeloTTS",
            "type": "TTS",
            "developer": "MyShell.ai / MIT",
            "license": "MIT",
            "commercial": "permissive",
            "size": "~50–200 MB per language",
            "languages": "EN (multi-accent), FR, ES, ZH, JP, KR",
            "quality": "Solid mid-tier — better than Piper, below Kokoro/XTTS",
            "speed": "CPU real-time inference is an explicit design goal",
            "streaming": "partial",
            "offline": "yes",
            "min_hw": "CPU-only (desktop)",
            "pi": false,
            "best_for": "Permissive multilingual (incl. French) CPU realtime TTS",
            "grade": "A",
            "notes": "One of the few MIT models with a proper French voice. Quiet since 2024 but widely deployed."
        },
        {
            "model": "StyleTTS 2",
            "type": "TTS",
            "developer": "Yinghao Aaron Li et al.",
            "license": "MIT",
            "commercial": "permissive",
            "size": "~150–350 M",
            "languages": "English-only (official)",
            "quality": "Paper reports surpassing human-recording MOS on LJSpeech; the architecture behind Kokoro",
            "speed": "Faster than realtime on GPU; heavy on CPU (unverified)",
            "streaming": "no",
            "offline": "yes",
            "min_hw": "~4 GB VRAM",
            "pi": false,
            "best_for": "Research-grade naturalness; fine-tuning your own high-quality English voice",
            "grade": "B",
            "notes": "MIT, but repo asks users not to clone voices without consent. Kokoro is its production-ready descendant."
        },
        {
            "model": "Bark",
            "type": "TTS",
            "developer": "Suno",
            "license": "MIT",
            "commercial": "permissive",
            "size": "~1 B (small variant available)",
            "languages": "13+ (incl. FR)",
            "quality": "Expressive text-to-AUDIO (laughter, sfx) but unstable/hallucination-prone; not precision TTS",
            "speed": "Below realtime on CPU; needs GPU (~12 GB full, ~8 GB small)",
            "streaming": "no",
            "offline": "yes",
            "min_hw": "8 GB VRAM (small) / 12 GB (full)",
            "pi": false,
            "best_for": "Expressive one-off audio with nonverbal sounds",
            "grade": "A",
            "notes": "Relicensed MIT in 2023. No reliable voice control; unmaintained since ~2023."
        },
        {
            "model": "Orpheus TTS (3B/1B/…)",
            "type": "TTS",
            "developer": "Canopy Labs",
            "license": "Apache-2.0",
            "commercial": "permissive",
            "size": "3 B flagship; 1B/400M/150M variants",
            "languages": "English primary; multilingual research checkpoints",
            "quality": "Human-like emotive speech from a Llama-3.2 backbone; strong zero-shot cloning",
            "speed": "~200 ms streaming latency on GPU; runs in GGUF via llama.cpp-style runtimes",
            "streaming": "yes",
            "offline": "yes",
            "min_hw": "~6–8 GB VRAM for 3B quantized",
            "pi": false,
            "best_for": "Streaming conversational TTS inside an LLM-style local stack (same tooling as your LLMs)",
            "grade": "B",
            "notes": "Llama-architecture → slots into existing GGUF/llama.cpp pipelines. Attractive if you already run local LLMs."
        },
        {
            "model": "Parler-TTS (mini/large)",
            "type": "TTS",
            "developer": "Hugging Face",
            "license": "Apache-2.0",
            "commercial": "permissive",
            "size": "880 M (mini) / 2.3 B (large)",
            "languages": "English-only",
            "quality": "Good naturalness with text-description voice control ('a calm female speaker…')",
            "speed": "~realtime on mid-range GPUs; slow on CPU (unverified)",
            "streaming": "partial",
            "offline": "yes",
            "min_hw": "~6 GB VRAM",
            "pi": false,
            "best_for": "Prompt-described voice characteristics without a reference clip",
            "grade": "B",
            "notes": "Effectively dormant since 2024; superseded by Kokoro/Chatterbox for most uses."
        },
        {
            "model": "XTTS-v2",
            "type": "TTS",
            "developer": "Coqui (defunct) / community fork",
            "license": "CPML-1.0 (Coqui Public Model License)",
            "commercial": "non-commercial",
            "size": "~470 M",
            "languages": "17 (incl. FR)",
            "quality": "Still one of the best open zero-shot voice cloners (6 s reference); natural prosody",
            "speed": "<200 ms streaming on GPU; CPU well below realtime",
            "streaming": "yes",
            "offline": "yes",
            "min_hw": "~4–6 GB VRAM",
            "pi": false,
            "best_for": "Hobby/personal multilingual cloning where non-commercial is acceptable",
            "grade": "A",
            "notes": "🔴 CPML = NON-COMMERCIAL (covers the audio output too), and with Coqui shut down there is NO ONE to buy a commercial license from — a dead end for any commercial product. Library code (MPL-2.0) is fine; the WEIGHTS are locked."
        },
        {
            "model": "Fish Speech / OpenAudio S1-mini",
            "type": "TTS",
            "developer": "Fish Audio",
            "license": "CC-BY-NC-SA-4.0 (weights); code Apache",
            "commercial": "non-commercial",
            "size": "0.5 B (mini); full S1 API-only",
            "languages": "multilingual (13+, incl. FR)",
            "quality": "Near-SOTA open cloning with emotion markers; S1 leads community arenas",
            "speed": "~realtime on RTX 4060-class (community, unverified)",
            "streaming": "partial",
            "offline": "yes",
            "min_hw": "~4–6 GB VRAM",
            "pi": false,
            "best_for": "High-quality zero-shot cloning for personal/research use only",
            "grade": "A",
            "notes": "🔴 Non-commercial share-alike weights — same commercial dead-end as XTTS-v2 (Fish sells API access separately)."
        }
    ]
}