Audio & Speech

Today's Stories

1. 1

   Anthropic clashes with Trump's White House, rejects Washington playbook

   The Trump administration has twice taken actions against Anthropic—labeling it a "supply chain risk" in April after the company refused Pentagon contract language, and imposing export controls on its Mythos and Fable AI models two weeks ago following discovery of a jailbreak. OpenAI, by contrast, announced it was withholding release of GPT-5.6 at the U.S. government's request on the same day those controls were relaxed. Anthropic, valued at $965 billion（約150兆円） and preparing for an IPO expected in the coming months, has refused the flattery, donations, and appointment of Trump allies that other tech giants (Meta, Amazon, Apple) and OpenAI have used to stay in the administration's favor. Trump administration officials have publicly attacked CEO Dario Amodei as a "liar" with a "God-complex" and an "ideological lunatic," and accused the company of "regulatory capture." Continued hostility could make it harder to sell public market investors on the stock listing and significantly hamper the company's ability to develop advanced AI models.

   Anthropic CEO Amodei reportedly called Trump "a feudal warlord" in a now-deleted Facebook post, and his sister and cofounder Daniela Amodei donated to Kamala Harris's campaign. Unlike OpenAI's policy chief Chris Lehane and cofounder Greg Brockman (the largest donor to Trump Super PAC MAGA Inc.), Anthropic has made no similar hires of Trump-aligned figures to its leadership.

2. 2

   Netflix uses AI-generated Gene Wilder voice for Wonka reality show

   Netflix is premiering Wonka's The Golden Ticket on September 23rd, a reality competition based on the fictional Wonka universe. The show's voiceover uses an AI-generated version of Gene Wilder's voice, created in partnership with AI audio company ElevenLabs and with consent from Wilder's family. This extends Netflix's pattern of using AI-generated celebrity voices for content—the company has previously recreated voices of Michael Caine and Stan Lee. For viewers, it means encountering synthetic versions of iconic figures in new productions, blurring the line between archival and synthetic media in mainstream entertainment.

   The two-part finale airs on September 30th. The show will feature 12 golden ticket winners and their chosen partners competing in a high-stakes social experiment, with one champion crowned by the end.

3. 3

   ML researcher seeks help debugging Pocket TTS implementation

   A machine learning researcher is attempting to implement Pocket TTS (a text-to-speech model from kyutai-labs) from scratch for learning purposes, since the authors have not released training or fine-tuning code. The researcher trained a single-speaker version on LJSpeech and LibriSpeech datasets but is encountering significant inference failures. This illustrates a common frustration in the open ML community: when researchers publish papers without releasing full training code, others must reverse-engineer the implementation from scratch, which is time-consuming and error-prone. For anyone trying to build or adapt published models, closed code can mean months of debugging.

   The researcher reports that despite low loss values during training (flow matching loss around 0.20 mse, very low EOS loss), the model barely generates meaningful speech even on text from its training set, and standard techniques like scheduled sampling and noise injection have not resolved the issue. This suggests a potential mismatch between training and inference that requires further investigation.

4. 4

   Ask HN: Best AI Voice Tool Beyond ElevenLabs for Character Acting?

   A Hacker News user asked for recommendations on AI voice narration tools that can handle emotional character acting better than ElevenLabs, which they found produced flat and inconsistent voices. The user has already tried Stable Diffusion 2.0's voice reference (limited to 15 seconds) and paid $75 for human voice acting that proved unusable. For creators producing animations or other voice-heavy content, the current state of AI voice tools may not yet deliver the emotional range and consistency needed to replace or supplement human actors—a gap that's driving users back to paid human labor, even when the results disappoint.

   The question remains unanswered in the body; no commenter has yet proposed a tool that solves this problem, suggesting the search for production-ready emotional AI voice acting is still unresolved.

5. 5

   NagaTranslate: AI translation and voice tool for India's low-resource Naga languages

   A developer has built NagaTranslate, a translation and speech pipeline for low-resource languages spoken in Nagaland, India, currently supporting Nagamese, Ao, and Sema. The project uses a commercial LLM API with optimized prompts and few-shot examples for translation, after initially experimenting with a fine-tuned NLLB (No Language Left Behind) model. Nagamese and other native Naga languages were primarily oral with very little standard parallel data, making this a low-resource NLP challenge. Building usable translation and voice tools for languages with scarce training data could help preserve and digitize these languages and make them more accessible.

   The project is seeking feedback on its architecture and strategies for improving the pipeline under strict resource constraints. The developer is sharing technical details and inviting community input on how to handle the unique challenges of working with oral languages that lack extensive written datasets.

6. 6

   Which AI Voice Agent Stack Has the Lowest Latency?

   Which AI Voice Agent Stack Has the Lowest Latency?