Speech: TTS and ASR
- Speech: TTS and ASRModel
Qwen3-TTS 1.7B CustomVoice open-source voice cloning model on Hugging Face
A post describes an open-source 1.7B text-to-speech model for voice cloning, with controls for emotion, age, tone, and accent, and multilingual support. The linked Hugging Face model card is the primary source; the author's claims about realism and latency are unverified.
Engineers evaluating self-hosted TTS can check the model card for licensing, languages, and controls.
- 1 month ago @multimodalart
Single-voice TTS model distilled from Kokoro-82M at 9MB
A post reports a single-voice, single-language text-to-speech model with about 9MB of weights, distilled from the 8M-parameter Kokoro-82M. The author says it runs anywhere and is very fast.
A tiny distilled TTS model could let engineers add fast local speech output on constrained devices.
- 1 month ago @KaitlynZhou
Voice clones of non-native English speakers rated more humanlike than originals
A NeurIPS 2026 paper cloned voices of 86 non-native English speakers using ElevenLabs, Coqui-XTTS and ChatterBox. In 4,000 human ratings, the clones were judged warmer, more authoritative and more humanlike.
It gives engineers evidence on how current TTS voice-cloning systems are perceived by listeners, including trust effects.
- Speech: TTS and ASRArticle
Multilingual speech-to-text benchmark for transcribe.cpp models on a laptop
A site showing accuracy and speed of speech-to-text models that transcribe.cpp supports, measured per language on FLEURS. Performance data comes from testing on a laptop CPU/GPU (4750U).
Lets engineers compare local ASR models on accuracy versus speed across languages before choosing one.
- 1 month ago @HuggingPapers
Google Releases a 4B Gemma Model for Speaker Diarization
Google released a 4B Gemma model fine-tuned to correct speaker labels and reduce word-level diarization errors across four benchmarks.
It may help engineers improve speaker labels in speech-processing pipelines.
- Speech: TTS and ASRPaper
Symbiotic Architecture Adds Audio Understanding to Frozen LLMs
The paper proposes an injector that writes audio-conditioned vectors into a frozen LLM’s KV cache, adding audio-understanding capabilities without fine-tuning the model’s weights.
Engineers can explore an approach to extending LLMs for audio understanding without changing their weights.
- 1 month ago @NVIDIAAI
NVIDIA releases Nemotron 3 Diarization model
Nemotron 3 Diarization tracks who spoke when in overlapping speech, supports up to eight speakers, and has 100M parameters. NVIDIA says it is available on Hugging Face.
Engineers can evaluate a speaker diarization model for handling overlapping voices.
- Speech: TTS and ASRModel
Youdao releases streaming ASR and translation models
The post reviews Youdao’s Confucius4-R2T2 for real-time ASR and Confucius4-T3PO for real-time translation. It says both support streaming and can be connected in sequence.
Engineers can explore a streaming ASR-to-translation setup for live speech applications.
- 1 month ago @kyutai_labs
Voice of Reason answers spoken math problems in speech
Kyutai Labs says Voice of Reason takes spoken math problems without transcription or a text LLM in the loop, then reasons and answers in speech. The post reports 77.1% on GSM8K, compared with 27.3% for GLM-4-Voice.
It describes a speech-native approach to spoken math reasoning without an intermediate transcription or text LLM.
- Speech: TTS and ASRProject
GrainSpeech: compact speech synthesis with less context
GrainSpeech is a compact speech synthesis project described as using less context for more detail. Its official repository is linked alongside a project page and paper.
Engineers can review the repository and paper for details on a compact TTS approach.
- 1 month ago @kadirnardev
Supersonic TTS Runs on CPU and Supports 31 Languages
The author says Supersonic runs on CPU and supports 31 languages. They plan to use it to generate synthetic data.
CPU support and multilingual coverage may make it useful to evaluate for local TTS and synthetic-data workflows.
- Speech: TTS and ASRProject
Xiaomi’s CocktailASR-1 targets a selected speaker in mixed audio
CocktailASR-1 is an end-to-end target-speaker speech recognition model that uses reference voiceprints to transcribe a selected voice in overlapping audio. The post says it supports English and Chinese.
It may help engineers evaluate speaker-conditioned ASR for crowded or overlapping audio.
- Speech: TTS and ASRProject
RVC WebUI for Training Voice Conversion Models
RVC is an open-source WebUI project for training voice conversion models. The repository describes training a model with 10 minutes or less of voice data.
Engineers can inspect a self-hostable voice conversion workflow and its model-training requirements.
Xiaomi-CocktailASR-1 transcribes a target speaker in mixed speech
A Hugging Face Space for transcribing only a target speaker in mixed speech, using an audio clip of that speaker to isolate their voice.
Engineers can try a speaker-conditioned ASR workflow on overlapping speech.
- Speech: TTS and ASRProject
OmniVoice: voice-cloning TTS for 600+ languages
OmniVoice is a Python voice-cloning text-to-speech project described as supporting 600+ languages. The post claims zero-shot synthesis and voice controls such as age, pitch, accent, and whispering.
Engineers can evaluate a multilingual TTS project for voice cloning and configurable speech generation.
- Speech: TTS and ASRPaper
Phoenix-VAD: Streaming Semantic Endpoint Detection
The paper introduces Phoenix-VAD, an LLM-based model for streaming semantic endpoint detection in full-duplex speech interaction. It uses a sliding-window training strategy.
Engineers building spoken dialogue systems can evaluate a semantic endpoint detection approach for streaming interactions.
- 2 months ago @ArtificialAnlys
Microsoft releases MAI-Transcribe-2 speech recognition model
Microsoft’s MAI-Transcribe-2 is a speech transcription model reported at 2.0% AA-WER and about 411× real time. It supports 60 languages, speaker diarization, word-level timestamps, and clean or verbatim transcription styles.
Engineers can compare its reported accuracy, speed, and price when evaluating speech recognition services.
- 2 months ago @realmrfakename
VibeVoice ASR Streaming combines transcription and diarization
The post announces the release of VibeVoice ASR Streaming, describing it as a streaming model that supports transcription and diarization in one model.
Engineers evaluating speech-recognition systems may want to assess its streaming and diarization capabilities.
- Speech: TTS and ASRArticle
Photon 2.1 adds streaming speech recognition
Photon 2.1 adds automatic speech recognition for Whisper, Qwen3-ASR, and Parakeet. Moondream reports that it beat each model’s reference engine in every matched test and supports NVIDIA B200.
Engineers evaluating self-hosted ASR can compare Photon’s reported performance across models and GPU types.
- 2 months ago @yoitsmanan
Phonon-1 speech recognition model released
Phonon-1 is a 782-million-parameter speech recognition model with a 415 MB download and Apache 2.0 licensing. Its post claims it transcribes an hour of audio in about two minutes on a MacBook Air.
Its reported size, license, and local runtime may be useful when evaluating self-hosted ASR models.
arxiv.org
