Modèles d’IA vocale locale (STT / TTS)
Réponse rapide
Si vous construisez un outil vocal privé et hors ligne, le modèle que vous pouvez réellement LIVRER est décidé par sa licence avant même de juger sa qualité. En reconnaissance vocale, vous êtes gâté : toute la famille Whisper (whisper.cpp et faster-whisper, tous deux MIT) est sûre pour le commercial, tourne hors ligne, et whisper.cpp descend jusqu'au Raspberry Pi — c'est pourquoi un produit local de voix-vers-texte comme DCENT_Voice livre exactement cela. Le Parakeet de NVIDIA (CC-BY-4.0) bat Whisper en précision-par-watt en anglais, et le STT en streaming de Kyutai est nativement anglais-plus-français — la paire linguistique canadienne exacte. En synthèse vocale, c'est un champ de mines : les meilleurs cloneurs, XTTS-v2 (le CPML non-commercial de Coqui, sans plus personne à qui acheter une licence) et Fish Speech (CC-BY-NC-SA), sont des impasses commerciales — tandis que Kokoro-82M et Chatterbox (tous deux permissifs) offrent une qualité sûre pour le commercial, et Piper tourne sur un Pi.
Triez par licence d'abord, qualité ensuite. Voix locale sûre pour le commercial aujourd'hui : whisper.cpp / faster-whisper / Whisper (MIT) et Parakeet (CC-BY) pour le STT ; Kokoro-82M, Chatterbox (MIT), MeloTTS et Piper (GPL — correct en auto-hébergé) pour le TTS. Traitez XTTS-v2 et Fish Speech comme à usage personnel seulement, et vérifiez toujours le checkpoint exact — Canary et Voxtral ont des scissions de licence par variante.
Licence : permissive MIT/Apache/BSD — commercial OK attribution CC-BY — commercial OK, crédit requis copyleft GPL — OK en auto-hébergé non-commercial impasse produit · ▪ Pi tourne sur du matériel de classe Raspberry Pi
| Modèle | Type | Licence | Langues | Vitesse / streaming | Matériel min. | Idéal pour |
|---|---|---|---|---|---|---|
| Reconnaissance vocale (STT) | ||||||
| whisper.cpp ▪ PiGeorgi Gerganov (ggml-org) | STT | permissiveMIT | multilingual (~99, incl. FR) | Faster than realtime on ordinary laptop CPUs (base/small); tiny/base run on Raspberry Pi flux : partial | CPU-only, down to Pi / phones | Shipping Whisper inside a product: one C/C++ binary, no Python |
| Same accuracy as the Whisper checkpoint it runs (minor loss at q4/q5) The obvious DCENT_Voice engine. `stream` example is sliding-window pseudo-streaming, not true streaming. | ||||||
| faster-whisperSYSTRAN (CTranslate2) | STT | permissiveMIT | multilingual (incl. FR) | ~4× faster than openai/whisper, less memory; faster-than-realtime large-v3 on mid-range NVIDIA flux : no | CPU workable (small/int8); 4–8 GB VRAM for large-v3 | Fastest full-accuracy Whisper on a GPU or beefy CPU |
| Same WER as reference Whisper for the same checkpoint The default server-side runtime; powers most self-hosted transcription stacks. | ||||||
| Whisper (tiny→large-v3) ▪ PiOpenAI | STT | permissiveMIT | multilingual (~99, incl. FR) | Original PyTorch slow; large-v3-turbo ~8× faster than large-v3 flux : no | CPU for tiny/base; ~10 GB VRAM for large-v3 fp16 | Multilingual accuracy baseline; the weights everything else runs or distils |
| large-v3 ~7.4% avg WER; the multilingual accuracy reference Weights + code MIT — fully commercial-safe. large-v3-turbo (809 M) is the practical default size. | ||||||
| Distil-Whisper (large-v3)Hugging Face | STT | permissiveMIT | English-only | ~6× faster than large-v3; drop-in with faster-whisper/whisper.cpp flux : no | CPU workable quantized; ~4 GB VRAM | English-only pipelines wanting large-v3 accuracy at small-model speed |
| Within ~1% WER of large-v3 on out-of-distribution English English only — a real limit for a bilingual EN/FR audience. | ||||||
| Voxtral Mini 3BMistral AI | STT | permissiveApache-2.0 | multilingual (strong EN/FR/ES/DE) | GPU-class inference; not an edge model flux : no | ~10 GB VRAM (bf16); quantized lower | Speech UNDERSTANDING (transcribe + summarize + answer) fully locally |
| Mistral claims it beats Whisper large-v3 on transcription; also does audio Q&A/summarization July 2025. Do NOT conflate with the separate 2026 'Voxtral TTS' models reported CC-BY-NC (unverified). | ||||||
| Moonshine (tiny/base) ▪ PiUseful Sensors / Moonshine AI | STT | permissiveMIT | English-only | ~5× faster than Whisper equivalents on short clips; runs on Pi-class edge flux : yes | CPU-only, Raspberry Pi / edge | Sub-second-latency voice commands and live captions on tiny hardware |
| Beats Whisper tiny/base WER at comparable size; compute scales with clip length Moonshine v2 (2026) adds a streaming encoder. English-only is the tradeoff. | ||||||
| Vosk ▪ PiAlpha Cephei (Kaldi) | STT | permissiveApache-2.0 | multilingual (20+, incl. FR) | Real-time streaming on CPU incl. Pi and Android flux : yes | CPU-only, Pi / phones | True streaming with partial results on the weakest hardware; mature bindings |
| Worse WER than Whisper-class, but respectable for its size Pre-transformer; pick it for latency and footprint, not accuracy. | ||||||
| WhisperXMax Bain (m-bain) | STT | permissiveBSD-2-Clause | multilingual (alignment per-language) | ~70× realtime batched large-v2 on GPU flux : no | ~8 GB VRAM for full diarization | Word-level timestamps and speaker diarization (meetings, subtitles) |
| Whisper accuracy + accurate word-level timestamps via forced alignment Diarization uses pyannote models needing HF gated-terms acceptance (still free/offline after download). | ||||||
| NVIDIA Parakeet TDT 0.6B (v2/v3)NVIDIA (NeMo) | STT | attributionCC-BY-4.0 | v2 English; v3 25 European (incl. FR) + auto-detect | RTFx ~3380 batched on datacenter GPU; community CPU/GGUF ports faster-than-realtime (unverified) flux : partial | ~2–4 GB VRAM; CPU via community ports | Best English accuracy-per-watt for batch transcription |
| v2 topped Open ASR Leaderboard at 6.05% avg WER — beats Whisper large-v3 on English CC-BY-4.0 = commercial OK with attribution. Native runtime is NeMo (heavier than whisper.cpp). | ||||||
| NVIDIA Canary (1b-flash / 1b-v2)NVIDIA (NeMo) | STT | attributionCC-BY-4.0 (flash/v2); original canary-1b CC-BY-NC | flash EN/DE/FR/ES + translation; v2 25 European | >1000 RTFx batched on GPU flux : no | ~4 GB VRAM | Combined transcription + speech translation (EN↔FR) in one local model |
| SOTA multitask ASR + speech-translation at release 🔴 License VARIES BY VARIANT: original canary-1b = CC-BY-NC (non-commercial); flash + v2 = CC-BY-4.0. Check the exact checkpoint. | ||||||
| Kyutai STT (1b-en_fr / 2.6b-en)Kyutai Labs | STT | attributionCC-BY-4.0 (weights); code Apache/MIT | English + French (1B); English (2.6B) | Natively real-time streaming, 0.5 s delay; an H100 serves 400 streams flux : yes | GPU recommended; 1B on consumer GPUs / Apple Silicon (MLX) | True low-latency streaming dictation — and EN+FR is exactly the D-Central language pair |
| Competitive WER with word-level timestamps; built for streaming Delayed Streams Modeling (from Moshi). CC-BY-4.0 = commercial OK with attribution. | ||||||
| Synthèse vocale (TTS) | ||||||
| Kokoro-82Mhexgrad | TTS | permissiveApache-2.0 | primarily English (strong); 8 langs thinner (FR limited) | Faster than realtime on modern desktop CPUs; trivially realtime on any GPU flux : partial | CPU-only (desktop); ~1–3 GB RAM | Best permissive-licensed quality-per-FLOP — the default for commercial local narration |
| Punches far above its weight — topped community TTS rankings vs far larger models Apache-2.0 weights trained only on permissive audio. Fixed voice pack, no cloning. Pi 5 borderline (unverified). | ||||||
| Chatterbox / MultilingualResemble AI | TTS | permissiveMIT | English (base); Multilingual 23+ incl. FR | Realtime on mid-range NVIDIA (A10/3060); CPU slow (unverified) flux : partial | ~6 GB VRAM; CPU/ROCm via community | The best MIT-licensed zero-shot voice cloning — the commercial-safe XTTS replacement |
| Beat ElevenLabs in blind preference (~63%, vendor eval); emotion-exaggeration control Output carries Resemble's PerTh neural watermark baked in — transparency plus or caveat depending on use. | ||||||
| MeloTTSMyShell.ai / MIT | TTS | permissiveMIT | EN (multi-accent), FR, ES, ZH, JP, KR | CPU real-time inference is an explicit design goal flux : partial | CPU-only (desktop) | Permissive multilingual (incl. French) CPU realtime TTS |
| Solid mid-tier — better than Piper, below Kokoro/XTTS One of the few MIT models with a proper French voice. Quiet since 2024 but widely deployed. | ||||||
| StyleTTS 2Yinghao Aaron Li et al. | TTS | permissiveMIT | English-only (official) | Faster than realtime on GPU; heavy on CPU (unverified) flux : no | ~4 GB VRAM | Research-grade naturalness; fine-tuning your own high-quality English voice |
| Paper reports surpassing human-recording MOS on LJSpeech; the architecture behind Kokoro MIT, but repo asks users not to clone voices without consent. Kokoro is its production-ready descendant. | ||||||
| BarkSuno | TTS | permissiveMIT | 13+ (incl. FR) | Below realtime on CPU; needs GPU (~12 GB full, ~8 GB small) flux : no | 8 GB VRAM (small) / 12 GB (full) | Expressive one-off audio with nonverbal sounds |
| Expressive text-to-AUDIO (laughter, sfx) but unstable/hallucination-prone; not precision TTS Relicensed MIT in 2023. No reliable voice control; unmaintained since ~2023. | ||||||
| Orpheus TTS (3B/1B/…)Canopy Labs | TTS | permissiveApache-2.0 | English primary; multilingual research checkpoints | ~200 ms streaming latency on GPU; runs in GGUF via llama.cpp-style runtimes flux : yes | ~6–8 GB VRAM for 3B quantized | Streaming conversational TTS inside an LLM-style local stack (same tooling as your LLMs) |
| Human-like emotive speech from a Llama-3.2 backbone; strong zero-shot cloning Llama-architecture → slots into existing GGUF/llama.cpp pipelines. Attractive if you already run local LLMs. | ||||||
| Parler-TTS (mini/large)Hugging Face | TTS | permissiveApache-2.0 | English-only | ~realtime on mid-range GPUs; slow on CPU (unverified) flux : partial | ~6 GB VRAM | Prompt-described voice characteristics without a reference clip |
| Good naturalness with text-description voice control ('a calm female speaker…') Effectively dormant since 2024; superseded by Kokoro/Chatterbox for most uses. | ||||||
| Piper ▪ PiRhasspy → Open Home Foundation | TTS | copyleftMIT (archived) / GPL-3.0 (maintained piper1-gpl) | multilingual (30+, incl. FR) | Faster than realtime on Raspberry Pi 4; near-instant on desktop CPU flux : partial | CPU-only, Raspberry Pi | Instant offline speech on the weakest hardware; voice-assistant responses |
| Clearly synthetic but pleasant and highly intelligible; the embedded/assistant standard 🔴 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. | ||||||
| XTTS-v2Coqui (defunct) / community fork | TTS | non-commercialCPML-1.0 (Coqui Public Model License) | 17 (incl. FR) | <200 ms streaming on GPU; CPU well below realtime flux : yes | ~4–6 GB VRAM | Hobby/personal multilingual cloning where non-commercial is acceptable |
| Still one of the best open zero-shot voice cloners (6 s reference); natural prosody 🔴 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. | ||||||
| Fish Speech / OpenAudio S1-miniFish Audio | TTS | non-commercialCC-BY-NC-SA-4.0 (weights); code Apache | multilingual (13+, incl. FR) | ~realtime on RTX 4060-class (community, unverified) flux : partial | ~4–6 GB VRAM | High-quality zero-shot cloning for personal/research use only |
| Near-SOTA open cloning with emotion markers; S1 leads community arenas 🔴 Non-commercial share-alike weights — same commercial dead-end as XTTS-v2 (Fish sells API access separately). | ||||||
Données ouvertes (CC BY 4.0) : CSV · JSON · API : /wp-json/dc/v1/voice-ai-models?type=STT
Produits, réparations et guides connexes
- comment D-Central diagnostique les réparations ASIC
- bibliothèque de dépannage ASIC
- manuels ASIC et guides de réparation
- hashboards de remplacement
- cartes de contrôle ASIC
- blocs d’alimentation ASIC
- hashboard de remplacement pour la famille S19
- carte de contrôle de remplacement C52
- bloc d’alimentation APW12 pour S19
- comparer les specs dans la base de mineurs ASIC
- comparer les specs des mineurs ASIC
- base de mineurs ASIC
- services de réparation ASIC
- specs et rentabilité de l’Antminer S19
- acheter un Antminer S19 testé
- guide d’entretien Antminer S19
- service de réparation Antminer S19
- specs de l’Antminer S21
- Bitmain Antminer S21
- guide d’entretien Antminer S21
- puce BM1370BC pour S21 Pro
Dernière révision: 27 juillet 2026.