AlphaCephei

Voting, Ensembles and bringing AI to life alphacephei.com/nsh/2022/06/14…

Nice upsampler, includes 8khz input, very rare and important github.com/woongzip1/Univ…

Interesting approach to integration of reasoning into audio models

@CatGodSandHive Unfortunately we didn't get to the point to check pronunciation easily. We'll need some reliable GOP implementation, not yet exists.

@alphacep "Pronunciation issues" are a good test, because they don't just check for "smoothness." They highlight where models still stumble.

We systematically test modern TTS engines on Russian dataset. Qwen feels most interesting one. Good clarity and sound quality, reasonable intonation. Issues with pronunciation as always, it is a common thing. VibeVoice hallucinates. Fish is reasonable but a bit plain.

@FeitengLi Here is the updated table with 0.6B included. Also good actually, thanks for the suggestion.

@alphacep What about 'Qwen3 0.6B TTS'?

It would be nice to find English data for a similar test. Many people test CER but few signal quality (UTMOS or others) and intonation.

@rdesh26 "Comparable or greater accuracy" is simply not true. TDT is usually less accurate in noisy conditions because duration prediction fails. Only good for slow clean speech. Depth-first convolutions add to that.

A great write-up about the token-and-duration transducer, and why it makes NVIDIA's ASR models faster than anyone else!

@realmrfakename The real secret is that you don't to have to use network of size of 2M params, those are not going to be reliable in noise. A lightweight ASR of 20M params can do much better.

@alphacep What do you think of Silero VAD?

Finally a good VAD github.com/FireRedTeam/Fi… so many developers use bad but popular VAD these days

@ZDi____ Hope you know about sidon as well, it is an improved version

How did I not know this before? The guys over at Google proved that you can have a very good speech restoration model by using features extracted from a pretrained model.

What if you had nano-banana for audio? AudioChat is a multi-modal LM that performs fine-grained understanding, generation, and editing of multi-source scenes By diffusing continuous latents, it generates 48khz stereo edits with great input adherence: wanchichen.github.io/audiochat/

Text-to-speech models can’t get your address right? Turns out you’re not the only one. 📢New preprint! State-of-the-art speech models get 44% of street names wrong — and non-English primary speakers suffer twice the error impact!

🚀 The MOSS-TTS Family is here. From zero-shot cloning to real-time VoiceAgents, we have released our most powerful suite of audio models yet. The Lineup: MOSS-TTS Flagship: The industry's best zero-shot voice cloning. Features precise control over duration & Pinyin, capable of generating 1 hour of speech. MOSS-TTSD-v1.0: A new standard for dialogue generation. Comprehensive optimization for conversational scenes and small languages. Best-in-class performance in all evaluations. MOSS-VoiceGenerator: One-shot timbre generation. Create voices with a single sentence and complex instruction handling. MOSS-TTS-Realtime: Built for the next era of VoiceAgents. Synthesis starts in just 2 characters for instant response. MOSS-SoundEffect: Text-to-Audio sound effects to expand your creative toolkit. 🔥 Try it now: studio.mosi.cn/voice-synthesis 💻 Deploy (GitHub): github.com/OpenMOSS/MOSS-… 🔌 API Docs: studio.mosi.cn/docs/moss-tts Welcome to our demo. The era of 'childhood' for TTS is over. #MOSS #AI #TextToSpeech #TTS #OpenClaw #Agent #OpenMOSS #Opensource #VoiceAgent

Low-Resource ASR leaderboard from Microsoft huggingface.co/spaces/microso…

@realmrfakename @xeophon @modal @charles_irl Consider Hetzner 5x20TB it will be 200 euro per month.

@xeophon @modal @charles_irl got any suggestions for places to store ~100TiB on a tight budget?

well… guess it’s time to find 100TiB of storage elsewhere @modal @charles_irl

VibeVoice 7B finetuned to European languages better than the original huggingface.co/kugelaudio/kug…

@huseinzol05 UTMOS score is small for audio with clicks, noise corruptions like metallic noise. If your speech is clean and naturally sounding but phones sound different, your UTMOS will be good.

@alphacep i dont get about "smoothness", can u clarify

sorry for the delay sir

Google just released Waxal on Hugging Face 1,250 hours of transcribed speech in 14 African languages for ASR and 240 hours for TTS, covering over 100 million speakers huggingface.co/datasets/googl…

Active marketing in voice tech is a new thing of last year (our subreddit is full of dumb product mentions).