Speech: TTS and ASR
- 11 days ago @Prince_Canuma
NVIDIA RE-USE speech enhancement model now available on MLX-Audio
NVIDIA's RE-USE, a small multilingual universal speech enhancement model, is now available on MLX-Audio. The post's author says it helps clean up noisy recordings that speech-to-text models often struggle with.
A lightweight speech enhancement model on MLX could improve ASR input quality for noisy audio on local hardware.
- 12 days ago @victormustar
Ava: browser-based real-time voice assistant running on-device
A demo of Ava, a voice assistant that listens and responds in real time inside the browser, with all inference on the user's device. It combines LFM3-Omni for voice understanding and generation, Hugging Face WebGPU kernels, and TalkingHead with HeadAudio for avatar lip-sync.
Shows an open-source, fully on-device speech-to-speech stack running in WebGPU in the browser.
- Speech: TTS and ASRModel
NVIDIA releases open 9.6M-parameter RE-USE model for speech cleanup
NVIDIA researchers released RE-USE, an open multilingual audio-to-audio model with 9.61M parameters that reduces background noise, echo and other distortions to make speech clearer. A demo and model download are available on Hugging Face.
A small open audio enhancement model you can run yourself to clean speech before ASR or TTS pipelines.
- Speech: TTS and ASRModel
Ming 3 TTS open model reported to lead Artificial Analysis TTS ranking
The author says Ming 3 TTS, an open text-to-speech model hosted on Hugging Face by inclusionAI, ranks first among open TTS models and slightly beats Fish Audio Drama 3 according to Artificial Analysis.
A recent open TTS model to benchmark against Fish Audio Drama 3 if you self-host speech synthesis.
- 21 days ago @kimmonismus
Jack Dorsey speech model reportedly trained on one GB200 NVL72
A social post claims Jack Dorsey released a speech model that the author calls best in class, trained on a single NVIDIA GB200 NVL72 system. No model name or primary source is included.
If accurate, it shows how much speech model training compute can shrink; the claims are unverified.
- 21 days ago @RagavanThurair
Dessa: full-duplex speech-to-speech model for consumer GPUs
A post announces Dessa, a full-duplex speech-to-speech model that the author says was trained entirely on speech and can run on consumer GPUs. No primary source is linked.
Relevant for engineers evaluating self-hostable conversational voice models; the claims are unverified.
- 21 days ago @EricBezzam
Open TTS Leaderboard adds 31M and 35B models and UTMOS metric
The Open TTS Leaderboard added a 31M-parameter model (phonon by Gradium) and a 35B model (Drama-3 by FishAudio), reporting similar English WER (1.08 vs 1.19). It also added UTMOS, a metric for English output naturalness.
Shows how small and large open TTS models compare on WER and a new naturalness metric.
- Speech: TTS and ASRProject
Phonon 31M: open TTS model running on Apple Neural Engine in Swift
A Swift port of the open Phonon 31M text-to-speech model using CoreML on Apple's Neural Engine, with 103 voices and a text normalizer ported from Gradium. The author claims about 20 ms to first audio and about 20x real time on an M5 Pro.
Gives engineers a runnable, open-source on-device TTS option for Apple platforms, with code and model weights linked.
- Speech: TTS and ASRProject
Little Voices: on-device voice app using Phonon TTS and Pipecat on iPhone
An open-source iOS demo app where an angel and devil on your shoulder answer questions by voice. It runs Pipecat in embedded Python with Apple SpeechAnalyzer for speech-to-text and Gradium Phonon 31M for TTS, with no network inference.
A concrete example of running speech-to-text and TTS models fully on a phone, with a pipeline you can inspect.
- 24 days ago @_jakubchmura
Phonon: 31M-parameter open TTS running in the browser via WebGPU
A post reports porting Phonon, a 31M-parameter open text-to-speech model by GradiumAI, to run fully in the browser on WebGPU at half precision. The author cites about 47x real time on an M4 Pro and 7x on a Phone 13 Pro.
Shows a small open TTS model running client-side on consumer hardware, relevant to on-device voice features.
- 24 days ago @andimarafioti
Phonon: 40M-parameter multilingual TTS model released with speech-to-speech
Phonon is a text-to-speech release, per the author, with a roughly 40M-parameter model covering multiple languages and a library of voices. It can be used locally via the speech-to-speech package with a --tts phonon option.
A small multilingual TTS model you can run locally could suit low-latency or on-device voice pipelines.
- 24 days ago @neilzegh
Open-source TTS models half the size of Kokoro with voice cloning
The author announces open-source text-to-speech models about half the size of Kokoro (82M parameters) that score higher on naturalness and accuracy and support voice cloning. No link is provided.
Smaller self-hostable TTS models with voice cloning could lower the cost of running speech synthesis locally.
- 24 days ago @kyutai_labs
Kyutai releases smaller Pocket TTS models down to 1.9M parameters
Kyutai Labs says it released smaller Pocket TTS versions, from the original ~100M parameters down to an 18.2M model, a 6.9M model, and a 1.9M model. The author claims the 1.9M model runs 103x faster than real time on one CPU core, with voice cloning.
Offers small self-hostable TTS with voice cloning that can run on CPU, relevant for on-device speech apps.
- 24 days ago @TONGYI_SpeechAI
Streaming ASR: decoding speech without seeing future audio
A post from the Tongyi Speech AI account explains streaming automatic speech recognition, where the system must commit to words while the speaker is still talking. No linked page preview is provided.
Streaming ASR constrains latency and lookahead, which shapes model and decoder design for live voice apps.
- Speech: TTS and ASRProject
Streaming Drama-3 TTS added to huggingface speech-to-speech pipeline
A pull request adds Fish Audio's Drama-3 as a streaming text-to-speech option (--tts drama3) in Hugging Face's speech-to-speech repo, with optional voice cloning and expressive delivery such as whispering or laughing.
Gives engineers a self-hostable streaming TTS backend with voice cloning for building speech-to-speech pipelines.
- 25 days ago @AntLingAGI
Ant Ling open-sources five Ming-3.0 speech models across TTS, audio and ASR
Ant Ling announces open-source release of five Ming-3.0 models: TTS variants (0.5B voice design, 7.9B long-form), audio understanding and dialogue models, and an ASR model for dialects. No links are provided.
Open-weight speech models spanning TTS and ASR could be self-hosted for voice products, pending details on licenses and benchmarks.
- 25 days ago @TeksEdge
AuroraASR-1M: 1MB English speech recognition model that runs on CPU
AuroraASR-1M is a 949K-parameter Conformer-CTC speech recognizer (1.06MB INT8, Apache 2.0) trained on LibriSpeech. The author reports ~436x real time on one CPU thread, but 11.43% test-clean WER versus 5.6% for Whisper Tiny.
Useful for offline, CPU-only ASR where model size and latency matter more than accuracy.
- Speech: TTS and ASRProject
OpenVoice: open-source instant voice cloning for dubbing
The author describes using an AI agent with OpenVoice to dub a movie into Polish while keeping dialogue in sync on an M4 Mac mini with 16GB. OpenVoice is an instant voice cloning audio foundation model from MIT and MyShell.
Shows a local voice-cloning model being driven by an agent for practical video dubbing on consumer hardware.
- 1 month ago @ArtificialAnlys
HeyGen Voice tops Artificial Analysis Controlled Voice TTS Arena
Artificial Analysis reports HeyGen Voice ranks #1 on its Controlled Voice TTS Arena, with an Elo of 1,201 across 8 cloned voices in US and UK English, ahead of Qwen-Audio-3.1-TTS-Plus and ElevenLabs Eleven v4 Turbo.
Offers a third-party comparison of TTS quality, pronunciation, price per character and speed across commercial voice models.
- Speech: TTS and ASRModel
Whistle: 16.9 MB on-device speech-to-text model on Hugging Face
Cactus-Compute released Whistle, a 16.9 MB speech-to-text model hosted on Hugging Face. The author says it rivals Whisper base for on-device use.
A small ASR model that runs on-device could replace Whisper base where size and privacy matter.
