iFlytek has launched API access to SparkAudio 1.0, its speech foundation model, on the iFlytek Open Platform. The model understands speech end-to-end instead of the cascaded transcribe-then-reason pipeline, covering 99 languages and 202 dialects across transcription, translation, speaker diarization, emotion analysis and audio QA, and is trained entirely on domestic compute.
Speaker Diarization
说话人分离
Technique that segments and labels distinct speakers within multi-speaker audio.
Curated Intelligence
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.
NVIDIA released Nemotron 3.5 Transcribe, a speech LLM built on a hybrid Mamba-Transformer MoE architecture. It integrates a frozen diarizer with an ASR encoder to enable instruction-driven transcription, allowing users to control output formats (verbatim, written, translation) via text prompts. While it outperforms Qwen3-ASR in English benchmarks, it is offline-only and shows lower accuracy for Mandarin far-field meetings.
Nvidia released the open-source speech model Nemotron 3.5 Transcribe, which integrates transcription, translation, and speaker recognition for up to 40 languages. Utilizing a Mixture-of-Experts architecture (30B total parameters, 4B active), it runs on a single GPU and achieves top-tier benchmark results. The release also includes the Granary v2 dataset with 844,000 hours of audio and tools for domain-specific adaptation.
Nvidia released the open-weight Nemotron 3.5 transcription model, supporting 40 languages and translation to/from English for 31. With 30B parameters, it runs on a single chip and handles long recordings with speaker diarization, aiming to lower barriers for developers building meeting and medical transcription apps.
iFlytek released SparkAudio 1.0, a foundational speech model that unifies complex audio understanding, emotion recognition, and real-time speaker separation, overcoming information loss in traditional cascaded architectures. The upgraded X2.6 series includes a 4B on-device model achieving ~70% of cloud capabilities with low latency. With voice LLM API calls surging over 20x year-on-year, this marks a shift from simple text transcription to multi-source acoustic-semantic joint reasoning.
AI hardware startup Nirva announced a $25 million pre-Series A round, bringing its total funding to $33 million. Founded by ex-Meta Reality Labs employees, the company is launching a jewelry-like pendant that uses voice separation technology to listen only to the wearer, addressing privacy concerns in ambient AI devices.
iFlytek unveiled SparkAudio 1.0, a speech foundation model integrating machine translation, ASR, and semantic understanding. It supports real-time Chinese-English switching and speaker diarization, now deployed across iFlytek's input method and hardware ecosystem.
The Qwen team released Qwen3.8-LiveTranslate, reducing average latency (LAAL) in real-time interpretation from 2.8s to 2.3s via a new Interleave audio-text architecture. Building on support for 60 languages, the model introduces real-time speaker diarization, synchronized source-target output, and long-context disambiguation to enhance accuracy and voice cloning stability in multi-speaker scenarios.
Hugging Face Inference Endpoints demonstrates a solution combining ASR, diarization, and speculative decoding. The case aims to improve speech transcription efficiency and accuracy but lacks specific performance metrics or detailed reusable configuration guides.