Audio & Speech

Today's Stories

1. 1

   Paris AI voice startup Gradium raises $100M seed, expands to Bay Area

   Gradium, a Paris-based startup building voice AI models, closed its seed round at $100 million（約160億円） total, adding Nvidia to its investor roster after initially raising $70 million（約110億円） in December. The company is opening a Bay Area office to strengthen its talent position. Gradium competes in a crowded voice AI market alongside ElevenLabs (valued at $11 billion（約1.8兆円） in February) and major players like Google's Gemini. The startup has already landed major customers including French auto manufacturer Renault since launching in December, suggesting its technology for delivering voice at scale with ultra-low latency is gaining real commercial traction.

   The company is using fresh funding to establish itself near major AI labs in the Bay Area—a strategic bet that proximity to Anthropic, Google, Meta, and OpenAI is essential for competing in AI talent markets, even though Gradium's founders and initial backers were rooted in Paris's own AI hub.

2. 2

   Cohere releases open-source Arabic speech-to-text model

   Cohere released Cohere Transcribe Arabic, a 2-billion-parameter open-source model for Arabic speech recognition. The model is available on Hugging Face and through the Cohere API under the Apache 2.0 license. According to Cohere, it is the most accurate open-source Arabic speech-to-text system available and outperforms Whisper Large V3 and the standard Cohere Transcribe model in benchmarks. It addresses Arabic's specific challenges—dialect variety, bilingual Arabic-English conversations, code-switching, and specialized vocabulary—which are difficult for general speech recognition systems to handle accurately.

   Human ratings on a 1–5 scale show Cohere Transcribe Arabic outperforms both Whisper Large V3 and the standard Cohere Transcribe model in overall quality, dialect faithfulness, and code-switching. The model is available now on Hugging Face and via the Cohere API.

3. 3

   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.

4. 4

   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.

5. 5

   Reddit user struggles to replicate Kyutai Labs speech model on smaller dataset

   A developer attempting to implement the Pocket TTS model from Kyutai Labs on a smaller dataset encountered significant training and inference problems. Despite achieving low loss values during training (flow matching loss around 0.20 MSE, EOS loss at very small levels), the model failed to generate meaningful speech even on text from its training set at epoch 2800, and produced hallucinations and repeated phrases. The post highlights a gap between reported model performance and reproducibility in practice—even when loss metrics look good during training, the resulting model may not work as expected. This matters for developers and researchers trying to adapt published models to their own data and compute constraints, as the original authors did not release training or fine-tuning code.

   The developer tried multiple techniques to address the failure—scheduled sampling to reduce exposure bias and adding Gaussian noise to ground truth data—but neither resolved the core problem. The underlying cause of the model's inability to generate coherent output despite low loss remains unclear from the post.

6. 6

   Ask HN: Best AI Voice Narration Tool Beyond ElevenLabs?

   A user seeking character voice acting for animation asked the Hacker News community whether there is an AI voice tool better than ElevenLabs at conveying emotional tone and character consistency, after finding ElevenLabs outputs felt flat and after spending $75 on a human voice actor whose work proved unusable. The question reflects a real gap in current AI voice technology — existing tools struggle with emotional expressiveness and character differentiation at scale, forcing creators to either accept poor results or pay for human labor that may not meet their needs.

   This is a community question seeking recommendations; the body does not announce a solution or availability date, only the user's frustration with existing options and limitations of prior tools (Stable Diffusion 2.0's 15-second voice reference limit).