Large language models get the headlines, but real AI products are built from several kinds of models. A voice agent needs speech-to-text, an LLM and text-to-speech. A RAG assistant needs an embedding model and a reranker. In 2026 open models in each of these categories became excellent — and most can run on modest hardware. Here are the ones to know.
Speech-to-text (automatic speech recognition)
What changed
Whisper made open speech recognition mainstream, but in 2026 newer models overtook it on accuracy. Hugging Face's Open ASR Leaderboard shows that models pairing a Conformer audio encoder with an LLM decoder are now the most accurate, while CTC and TDT decoders are 10 to 100 times faster (Hugging Face).
Leading open models
- Cohere Transcribe — a 2B Apache 2.0 model released in March 2026 that took first place on the Open ASR Leaderboard at launch with a 5.42% average word error rate, covering 14 languages (The Decoder).
- IBM Granite Speech 4.1 — a 2B Apache 2.0 model that IBM reports at 5.33% word error rate on the same leaderboard (IBM).
- NVIDIA Parakeet TDT 0.6B v3 — small and very fast, covering 25 European languages with word-level timestamps and up to three hours of audio, under CC-BY-4.0, which allows commercial use (model card).
- Whisper large-v3 and v3-turbo — still the broadest language coverage (99 languages) and strong on long recordings; a safe default for multilingual transcription.
- Mistral Voxtral — Voxtral Mini Transcribe Realtime and Voxtral Small are Apache 2.0 for speech understanding (Mistral docs).
How to choose: Whisper for the widest language coverage, Parakeet for speed and European languages, Cohere Transcribe or Granite Speech for top English accuracy.
Text-to-speech (TTS)
- Qwen3-TTS — 0.6B and 1.7B models under Apache 2.0, released in January 2026, with voice cloning from about three seconds of audio across 10 languages (Simon Willison).
- Chatterbox (Resemble AI) — MIT-licensed models from a tiny CPU-friendly version to a 500M multilingual model covering more than 23 languages; every output carries a built-in watermark (GitHub).
- Kokoro-82M — a remarkably small Apache 2.0 model with 54 voices in eight languages, ideal for fast, cheap narration where voice cloning is not needed (model card).
Licence warnings: Mistral's Voxtral TTS is CC BY-NC 4.0, which forbids commercial use, and Microsoft withdrew the code for its VibeVoice TTS model in 2025 after misuse. Always check terms before shipping a voice product — and get consent before cloning anyone's voice.
Embeddings and rerankers (for RAG and search)
Embedding models turn text into vectors so that search can find passages by meaning. They are the heart of every RAG system.
- Qwen3-Embedding (0.6B to 8B, Apache 2.0) — the 8B model ranked first on the MTEB multilingual leaderboard when released, supports 32K-token inputs and adjustable vector sizes, and comes with matching Qwen3-Reranker models (model card).
- EmbeddingGemma 2 — a 740M model from Google released in October 2026, suited to on-device and low-latency search (Gemma releases).
- IBM Granite Embedding Multilingual R2 — Apache 2.0 embeddings covering more than 200 languages, including a tiny 97M model (IBM).
Check the live MTEB leaderboard when choosing — rankings change often — and remember that some popular embedding models carry non-commercial licences.
Tip: a reranker often improves RAG quality more than switching to a bigger embedding model. Retrieve 50 candidates with embeddings plus keyword search, then let a reranker pick the best 5–10.
Putting it together
A self-hosted voice agent
| Step | Open model options |
|---|---|
| Speech-to-text | Parakeet, Cohere Transcribe, Whisper |
| Reasoning | Gemma 4, Qwen3.8-27B, Muse Glimmer |
| Text-to-speech | Qwen3-TTS, Chatterbox, Kokoro |
Everything stays on your servers — useful for call centres, clinics and banks — with the language model chosen from the best open-source LLMs.
A private RAG assistant
| Step | Open model options |
|---|---|
| Document parsing | PaddleOCR-VL, Docling (OCR guide) |
| Embeddings | Qwen3-Embedding, EmbeddingGemma 2, Granite Embedding |
| Reranking | Qwen3-Reranker |
| Answering | Gemma 4, Qwen3.8, DeepSeek V4.1-Flash |
See our self-hosting guide for the infrastructure side.
How Zinetics can help
We build voice agents, transcription pipelines and private RAG assistants — choosing open or commercial models for each step based on accuracy, latency, privacy and licence. Explore our AI services or tell us what you want to build.