Live intelligence

Latest AI updates

11 updates
Models

ElevenLabs releases Scribe v3 and Scribe v3 Live speech recognition models

ElevenLabs has released Scribe v3 and Scribe v3 Live, speech recognition models designed to transcribe accurately amid background noise and crosstalk. Artificial Analysis benchmarks show Scribe v3 narrowly beating OpenAI, Google, and Microsoft on accuracy. The tradeoff is speed and cost: rivals such as Microsoft and xAI run many times faster and charge roughly half the price. Scribe v3 Live uses conversational context to help voice agents avoid cutting off callers mid-sentence.

ElevenLabsspeech recognitionvoice agents
Research

Swanada: Split-Encoder Capability-Routed Conformer CTC and Sub-30 MB On-Device Speech Recognition Across Twelve Scripts

Swanada is an open-source 16 kHz streaming ASR system covering Classical Sanskrit, Indian English, and ten Indian languages across browser WebAssembly, Node.js, and Python. It uses capability-routed shared INT8 Conformer encoders: a 17-layer 177.38 MiB Studio tier at d=512 and a distilled 14-layer 25.35 MiB Mobile tier at d=256, paired with twelve hot-swappable 515.0 KiB script-specific CTC heads. The design reports 11.6x and 67.8x reductions in 12-language storage, with runtime techniques including dynamic positional synthesis, MatMul pointwise convolutions, and QInt8 head calibration. On held-out and open public test clips, it reports 5.43% and 8.87% CER on Studio, 18.56% and 19.23% on Mobile, at 0.0097x-0.0264x RTF with 30.6 ms median TTFT.

Speech RecognitionOn-device AIMultilingual ASR
Industry

Egyptian and Nigerian English join the Open ASR Leaderboard

The Open ASR Leaderboard now includes Egyptian and Nigerian English, expanding its coverage of regional English accents. This addition enables more accurate evaluation of speech recognition models on these dialects. The update reflects a growing focus on diverse language varieties in ASR benchmarking.

Open ASR Leaderboardspeech recognitionEnglish dialects
Products

Soniox STT v6 Adds Agent Text Input and Echo Filtering for Voice Agents

Soniox has released STT v6, a speech-to-text API update that introduces agent text input, allowing the model to read what the voice agent says while transcribing the caller, and model-level echo filtering to remove leaked agent audio. The update is available as stt-rt-v6 for streaming and stt-async-v6 for files, supporting 60+ languages, with pricing at $0.12/hour streaming and $0.10/hour async. No v6 benchmark numbers are published yet; v5 scored 1.34% semantic WER on Pipecat, and v5 models will be deprecated on December 15, 2026.

Soniox STT v6voice agent technologyspeech recognition
Research Editor selected

LipLingo: CNN Model for Lip Reading

LipLingo is a robust lip-reading system built on the LipNet architecture, incorporating a standardized preprocessing pipeline, rotational validation for speaker generalization, and flexible similarity evaluation for sentence-level assessment. Using 3D convolutional and bidirectional recurrent layers with CTC loss, it achieves 95.0% character accuracy and 87.2% sentence accuracy on the GRID dataset, outperforming the baseline LipNet. The work offers insights into improving lip-reading robustness for real-world deployment.

Why it matters: Presents a concrete, reproducible improvement in lip-reading accuracy with clear benchmarks, valuable for researchers and developers in speech recognition and assistive technology.

lip readingdeep learningspeech recognition
Models Editor selected

NVIDIA Releases Nemotron 3.5 Transcribe Open Speech LLM

NVIDIA has released Nemotron 3.5 Transcribe, an open-weights speech LLM that handles 40 languages and recordings up to one hour from a single model. The hybrid Mamba-Transformer MoE has roughly 34B total and 4.3B active parameters, with built-in speaker attribution, word timestamps and prompt-selected output styles. NVIDIA reports 4.71% mean WER on nine English Open ASR sets, the best open-weights score, though commercial APIs still lead. Weights ship in BF16, FP8, NVFP4 and GGUF under a commercial-use license, with NVFP4 fitting an RTX 5090.

Why it matters: Developers get a concrete open, commercially licensed speech model that combines transcription, diarization, timestamps and translation, with deployment details and benchmark context.

NVIDIANemotronspeech recognition
Models Editor selected

Nvidia releases open Nemotron 3.5 Transcribe for transcription, translation, and speaker recognition

Nvidia has released Nemotron 3.5 Transcribe, an open speech recognition model that handles transcription, speaker recognition, and translation and accepts instructions as text prompts. Nvidia says it leads several benchmarks while using only one-sixth of the active parameters of its strongest open rival. The release also includes a dataset of more than 800,000 hours of speech and tools for fine-tuning.

Why it matters: It gives readers a concrete new open speech model release with combined capabilities, benchmark claims, and supporting data and fine-tuning tools.

Nvidiaspeech recognitionopen models
Models

Nvidia releases open transcription model for recorded audio

Nvidia has released a free speech-to-text model capable of transcribing up to 40 languages. The model works on recordings rather than live audio, positioning it for meeting and medical transcription use cases. It leaves real-time transcription to competitors such as ElevenLabs.

Nvidiaspeech recognitionopen models
Models Editor selected

IBM Releases Granite Speech 5.0 2B, an Apache 2.0 Multilingual Speech Model

IBM released Granite Speech 5.0 2B, a 2B-parameter open speech-language model for transcription and translation in 6 languages under Apache 2.0. It combines a 16-block Conformer speech encoder with a window q-former and a fine-tuned granite-4.0-1b-base backbone, and adds language-ID prompting, keyword biasing up to 2,000 terms, and requested punctuation and capitalization. IBM reports 5.68% average English WER on the Open ASR Leaderboard, ranking second among open-weight models behind Qwen3-ASR-

Why it matters: Developers get a concrete, benchmarked open-weight ASR and translation release with deployment options and licensing details they can evaluate directly.

IBM Granitespeech recognitionopen-weight models
Tutorial

Independent CPU Benchmark Pits Cactus Whistle Against Whisper Base for Speech-to-Text

A MarkTechPost coding guide benchmarks Cactus Compute's 16.9 MB Whistle speech-to-text model against Whisper tiny and base on seeded LibriSpeech test-clean and test-other samples, using speaker-cluster bootstraps with Holm correction. The author reports a lead over greedy Whisper base on test-other and no significant gap against beam search, with a speed advantage that narrows on longer clips. The tutorial also documents a 30-second single-pass window where longer audio is silently truncated, pl

Cactus ComputeOpenAI Whisperspeech recognition
Products Editor selected

Mistral AI Releases Voxtral 3 Speech Models and LiveKit-Based Voice Agent SDK

Mistral AI launched Voxtral 3, a speech model family with three transcription models (Small, Mini and Mini Realtime) plus a text-to-speech model, all in public preview on the Mistral API. The release includes a voice agent SDK built on LiveKit that streams listening, LLM reasoning, tool calls and speech synthesis in one pipeline over SIP and WebRTC. Mistral reports 2.1% AA-WER for Small Transcribe 3 and 18.4% cpWER for diarization, while realtime accuracy lags Gemini 3.5 at low latency. Pricing

Why it matters: It gives AI builders concrete details on a new speech-to-text, TTS and voice-agent stack, including benchmarks, latency trade-offs and per-minute pricing that inform build-versus-buy decisions.

Mistral AIVoxtral 3speech recognition