Artificial Analysis reports that SOV-Transcribe-1 is now the top open weights speech-to-text model, achieving 2.3% on the AA-WER Index with 4.6B active parameters. It ranks 7th overall among 44 models including proprietary ones, released by Salesforce and the Missionforce Sovereign AI team.
Word Error Rate (WER)
词错误率
Ratio of erroneous words to reference words, the core metric for transcription quality.
Curated Intelligence
Moondream has released Parakeet Ultra, a post-trained version of the NVIDIA Parakeet TDT 0.6B v3 architecture, now available on OpenRouter. The model maintains the original speed and structure while achieving lower word error rates across benchmarks, supporting 25 European languages with segment and word timestamps at $0.06 per hour.
Neuralink released early feasibility data from its VOICE project, showing an ALS patient achieved 75 words per minute output within the first week of implantation, with a 5.8% word error rate on a large vocabulary. While this speed approaches normal conversation and highlights BCI potential for restoring communication, the results are based on a single participant and lack peer-reviewed validation.
NVIDIA released the open-weight speech model Nemotron 3.5 Transcribe, featuring an MoE architecture that supports 40 languages and precise transcription/diarization of up to 8 overlapping speakers. The model tops several open-source benchmarks and comes with the Granary v2 data pipeline, targeting use cases like meeting minutes and customer service.
Mistral AI launches the new Voxtral 3 speech model family, including Text-to-Speech and various Speech-to-Text variants. Small Transcribe 3 ranks #3 on the Artificial Analysis leaderboard, with significant improvements in real-time streaming accuracy and competitive TTS pricing.
Hugging Face, in collaboration with Appen and Dataocean, introduced a 'Benchmaxxer Repellant' mechanism to the Open ASR Leaderboard. By incorporating private, unpublicized test sets, this update aims to prevent models from overfitting or gaming public benchmarks, thereby enhancing the reliability of generalization metrics.
Alibaba's Qwen team released two new models: Qwen3-TTS-VD-Flash for voice design and Qwen3-TTS-VC-Flash for voice cloning. The former allows fine-grained control over timbre, emotion, and persona via natural language instructions, while the latter supports 3-second multi-language voice cloning. Benchmarks indicate superior performance over GPT-4o-mini-tts and ElevenLabs in InstructTTS-Eval and multilingual WER metrics.
Tongyi Lab released a major update to Qwen3-TTS, expanding support to 49 high-quality voices, 10 languages, and 9 Chinese dialects. The model achieves lower average Word Error Rate (WER) than ElevenLabs and GPT-4o-Audio-Preview on the MiniMax multilingual test set, with significantly improved prosody and human-like naturalness. It is now accessible via DashScope API and Hugging Face demos.
Hugging Face Transformers now supports n-gram language model boosted decoding for Wav2Vec2. By integrating external n-gram LMs with CTC outputs, the method reduces Word Error Rate (WER) by approximately 30% on the Common Voice dataset. This allows developers to significantly improve speech recognition accuracy without retraining the base model.