speech

"The new Gemini 3.1 Flash TTS allows developers to steer speech output using natural language instructions, integrating emotional nuance and pacing directly into the generation pipeline."

"2026年4月16日、Google Cloudから Gemini 3.1 Flash TTS のプレビュー版が公開されました。70を超える言語、30種類のプリセット音声、そして200以上の「オーディオタグ」 で囁き・叫び・笑い・ため息までテキストの中で自在に指示できるという、音声合成の世界をまた一段引き上げるモデルです。"

"Conventional evaluation protocols may therefore overestimate generalization and clinical utility, highlighting the need for strictly speaker-independent evaluation."

"Princess Afro Consulting has launched 'AI Kasuhara Guard', which uses AI to detect customer harassment from business partners and customers, leaving conversational text to ensure visibility and evidence of customer situations."

"The hard part isn't getting it to run once in a Jupyter Notebook; it's crafting that model into a reproducible, debuggable application that you can truly call your own."

"Since there wasn't one, I decided to make it, so I Fine-tuned a model based on Qwen3-ASR-1.7B using LoRA. The result is lilfugu."

"Experiments on the VoiceBank-DEMAND dataset demonstrate that GatherMOS consistently outperforms DNSMOS, VQScore, naive score averaging, and even learning-based models such as CNN-BLSTM and MOS-SSL when trained under limited labeled-data conditions."

"Feature analysis reveals that pitch variability and spectral richness (spectral centroid, bandwidth) are key discriminative cues."

"Across two studies (Interactive, N=24; Observational, N=19), we find that voice conversion increases trust and acceptability for benign responses and encourages perspective-taking, while automated analysis in search of quality-of-service disparities, reveals {accent x gender} disparities in alignment and verbosity across SpeechLLMs."

"The new model can generate natural-sounding, high-fidelity speech while allowing developers to control the emotion, pacing, and style of the voice through prompts, such as precisely adjusting tone, pauses, and emotional changes in narration or dialogue."

"Our newest audio model introduces granular audio tags that give you precise control to direct AI speech for expressive audio generation."

"Our newest audio model introduces granular audio tags that give you precise control to direct AI speech for expressive audio generation."

"Our method reduces bias word recognition errors by 16.3% compared to baseline systems, including on out-of-domain data."

"Google’s latest Realtime model Gemini 3.1 Flash Live audio removes that pipeline entirely. It processes audio natively. You stream audio in and the model streams audio back out."

"Our results indicate that the combination of web-scale unlabelled data and LLM-ensemble annotations is the most valuable for smaller models and low-resource languages."

"On LibriSpeech, the phoneme-based interface is competitive with the vanilla projector, and the BPE-phoneme interface yields further gains. On Tatar, the phoneme-based interface substantially outperforms the vanilla projector."

"Ladies and gentlemen, it is a great pleasure the confirm that llama.cpp (llama-server) now supports STT with Gemma-4 E2A and E4A models."

"ChatGPT speaks with a heavy accent, this is admittedly very hard to replicate"

"Neuralink enables nonverbal ALS patient to speak again with thoughts and AI-cloned voice"

"The proposed model achieved 97.8% accuracy and a macro F1-score of 0.98... highlight[ing] the potential of Transformer-based approaches in low-resource languages."

"Experiments on Mandarin and English benchmarks show that our method achieves competitive performance with state-of-the-art models using only 2.3B parameters, while also effectively mitigating hallucinations through our decoupling-oriented design."

"I used voicevox-openai-tts to wrap VOICEVOX as an OpenAI-compatible API, making it possible to generate Podcasts with easy-to-listen-to, high-quality Japanese voice."

"Our experiments suggest that the DAT-CFTNet leads to consistently improved performance over the existing models, including CFTNet and DCCRN, in terms of speech intelligibility and quality."

"The 2026 Multilingual Conversational Speech Language Model (MLC-SLM) Challenge has begun, aiming to further explore the potential of large language models in multilingual dialogue understanding, primarily involving acoustic and semantic information."

"In this article, I will explain the entire process of implementing the complete elimination of this hallucination by migrating from whisper-1 to gpt-4o-transcribe, accompanied by actual code."

"The high accuracy of MAI-Transcribe-1 is achieved through a separation architecture where Acoustic tokens handle acoustic characteristics... while Semantic tokens handle linguistic meaning structures... resulting in the ability to maintain low WER across 25 languages with a single model."

"The design of logs is critical; they must be saved at a granularity that allows for later analysis. Logs that cannot reconstruct 'what happened' after the fact hinder the improvement loop."

"OpenAI announced in April 2026 that the Realtime API has been upgraded to General Availability (GA), officially releasing the new model gpt-realtime. Compared to the traditional preview version, there are three major changes: SIP phone support, Remote MCP server support, and asynchronous function calling."

"発言の帰属が全話を通してほぼ正確でした。単に「話者A/話者B」ではなく、「イチロー：」「武豊：」と実名で正しく出力されており、この体験を技術的に解説したいと思います。"