10/15/2026
ElevenLabs Instant Soundtracks Scores Any Video
ElevenLabs Instant Soundtracks (Oct 15) writes an original score and drops timestamped SFX on any upload — live in ElevenCreative, ElevenMusic, Creative Engine, and the Studio API.
Anabel
66 Articles
Latest Contributions
10/15/2026
GPT-Live-1.1: ChatGPT Voice Gets 13 New Voices
OpenAI rolled out GPT-Live-1.1 and the smaller Quip to ChatGPT Voice: full-duplex Live mode with 13 new voices, better non-English accents, and less “checking…”. Free for all users today, API coming soon.
10/15/2026
Jais-ASR: Open Arabic Speech Recognition, 19 Dialects
MBZUAI’s IFM just opened Jais-ASR under Apache 2.0: one 4B-based model for Modern Standard Arabic plus 19 dialects, code-switching, optional diacritics, and a vLLM server out of the box.
10/15/2026
EmbeddingGemma 2 Spawns a Fast Open ASR Wave
Builders are turning Google’s open EmbeddingGemma 2 into on-device speech-to-text — QUAN-Transcribe for Vietnamese, norsk-asr for Norwegian, and a free 288× Mac app.
10/15/2026
Maya-2: Real-Time Indian TTS Hits Voice Arena #1
Maya Research launched Maya-2 for 11 Indian languages — #1 real-time Hindi on Voice Arena (Elo 1086), 75 ms first-byte claims, live API and Baseten. Open weights promised, not shipped yet.
10/14/2026
Grok Voice 3.0 Tops Speech-to-Speech Index
SpaceXSI's Grok Voice 3.0 is a closed speech-to-speech voice-agent model, #1 on Artificial Analysis' S2S Index at 86.9, live in the Voice Agent API today at $0.08/min.
10/14/2026
Speechmatics Oak Medical Multilingual STT
Speechmatics Oak is a closed medical STT that follows mid-conversation language switches and speakers, with 91% medical-term recall across 11 languages, 55+ langs, and API access today.
10/14/2026
QVAC TranscribePsy Open On-Device ASR
Tether AI Research’s QVAC TranscribePsy is a 254M open Apache-2.0 ASR for phones and laptops — 25 EU languages on laptop, 8 on phone — with day-0 Oct 14 weights, GGUF, and Core ML builds.
10/14/2026
Gnani Timbre v3.0 Adds Audio Tags for Indian TTS
Gnani AI’s Timbre v3.0 ships today with 13 inline audio tags, 13 voices across 10 Indian languages, and sub-200 ms first audio on the Vachana TTS API — expressive direction for agents and creators.
10/14/2026
Grok Imagine Quietly Adds In-App Music Gen
Creators are finding a Music panel inside Grok Imagine that turns text prompts into full songs and drops them onto your videos — with no official xAI post yet.
10/14/2026
Sarvam Bulbul V4 Flash Hits Sub-250ms TTS
Sarvam AI's new Bulbul V4 Flash TTS is built for Indian-language voice agents: under 250 ms to first audio, ₹20 per 10K characters (₹10 until Nov 12), and 83 voices across 11 languages on the same Bulbul API.
10/13/2026
Fish Audio Drama 3 Ships With Open Weights
Drama 3 has moved past preview. Fish Audio's directable TTS is now live in the app and API, with 35B-MoE open weights on Hugging Face and a Drama 3 option in ComfyUI's Fish Audio node.
10/12/2026
ASLP-lab SemBridge: Open TTS and Singing Voice Model
ASLP-lab released SemBridge on October 12, 2026 — open Apache-2.0 weights and inference for continuous-latent text-to-speech, note-conditioned singing, and mixed speech↔singing in one ~0.8B model you can run from Hugging Face today.
10/11/2026
LAION Humaneness Voice Converter: Open Chatterbox Voice Conversion That Keeps the Emotion
LAION released Humaneness Voice Converter on Oct 11, an open CC BY 4.0 speech-to-speech checkpoint built on Resemble AI's Chatterbox. It re-voices a source recording into a 5–10 second target voice while keeping more of the original emotion, with a runnable infer.py and a live comparison demo on Hugging Face.
10/9/2026
HeyGen Voice Debuts at #1 on Artificial Analysis TTS Arena, $15/M Chars Through Oct 31
HeyGen launched HeyGen Voice (heygen-voice-1), its first in-house text-to-speech model. It debuted at #1 on the Artificial Analysis Controlled Voice Arena with an Elo of 1,201, ahead of Qwen-Audio-3.1-TTS-Plus and Eleven v4 Turbo, and it's live in the HeyGen app and API at $15 per million characters through October 31.
10/8/2026
WorldSonus: Open-Weight Real-Time Stereo Sound for AI World Models
Noiz AI and HKUST released WorldSonus, an open-weight video-to-audio model that streams 48 kHz stereo sound in 100 ms chunks at 0.41 real-time factor on one H100, with prompts you can change mid-stream and camera-aligned stereo. Code is on GitHub and weights are on Hugging Face (CC BY-NC 4.0).
10/8/2026
TII's Falcon-ASR: A 1.6B Speech Model Built for Emirati Arabic
Abu Dhabi's TII launched Falcon-ASR, a compact 1.6B speech-to-text model for Emirati and Standard Arabic plus English, French, Spanish, and Portuguese, with word-level timestamps and a 20.92% average Arabic WER. Weights stay closed, but a free Hugging Face demo is live.
10/8/2026
Sesame's New Voice Model Lands on iOS and Android
Sesame's October release puts an updated Conversational Speech Model behind Maya, Miles, Simone, and Charlie, opens the app to everyone on iOS and Android, and gives each voice agent its own computer, MCP connectors, and a preview way to steer coding agents by voice.
10/7/2026
Artificial Analysis Music Arena v1.1: Suno v6 Leads Both Boards
Artificial Analysis rebuilt its AA-Music-Vocal and AA-Music-Instrumental leaderboards on 1,000 new prompts across 17 genres and 37,000+ blind panel votes. Suno v6 is #1 on both, Mureka V9.5 climbs 51 Elo on vocals, and MiniMax Music 3.0 leads open weights.
10/6/2026
Google Flow Music Spaces Now Export as VST3/AU Plugins
Google Flow Music now lets producers export Spaces—custom instruments and effects built from plain-language prompts—as VST3 and AU plugins that run inside their own DAW.
10/6/2026
Amazon Nova 2.5 Sonic: Smarter Voice Agents on Bedrock
Amazon’s new speech-to-speech model is GA on Bedrock with sharper reasoning, better tool calls, lower latency, and a 256K context—at the same price as Nova 2 Sonic.
10/4/2026
Ito TTS: Streaming Speech for a $5 ESP32-S3
Lokutor’s Ito packs 4.4M-param streaming TTS (~4.9 MB int8, 25 ms first chunk) aimed at a $5 ESP32-S3—bit-exact in QEMU/host today; physical board timings still unmeasured.
10/4/2026
Audio Sliders: 35 LoRAs for ACE-Step Music
takakhoo released Audio Sliders—35 rank-4 LoRAs that continuously steer frozen ACE-Step 1.5 XL turbo on mood, ensemble, energy, and more—MIT-licensed with a live demo.
10/4/2026
Whistle ASR: 16.9 MB On-Device Speech Model
Cactus Compute open-sourced Whistle—a 16.9 MB multilingual ASR with word timestamps and keyword biasing—plus a CPU Needle engine that claims ~11.1 ms TTFT on M4 Pro.