# Gradium

> Source: https://aiwiki.ai/wiki/gradium
> Updated: 2026-10-27
> Fact-checked: 2026-10-27
> Categories: AI Companies, Open Source AI, Speech & Audio AI, Voice AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Gradium." aiwiki.ai, 27 Oct 2026. https://aiwiki.ai/wiki/gradium
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Gradium** is a French [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) company, based in Paris, that builds real-time voice models and sells them through an API: streaming [text-to-speech](https://aiwiki.ai/wiki/text_to_speech_ai) (TTS), streaming [speech recognition](https://aiwiki.ai/wiki/speech_recognition) (speech-to-text, STT), [voice cloning](https://aiwiki.ai/wiki/voice_cloning), prompt-based voice design and live speech translation.[1][2] It was founded in 2025 by Neil Zeghidour, Olivier Teboul, Laurent Mazaré and Alexandre Défossez, researchers and engineers who previously worked at [Google DeepMind](https://aiwiki.ai/wiki/google_deepmind), [Google Brain](https://aiwiki.ai/wiki/google_brain), [Meta FAIR](https://aiwiki.ai/wiki/meta_fair) and Jane Street.[1][2] Three of them were on the founding scientific team of [Kyutai](https://aiwiki.ai/wiki/kyutai_labs), the Paris nonprofit research lab behind the speech model [Moshi](https://aiwiki.ai/wiki/moshi), and Gradium was the first company to spin out of that lab.[1][3][4][5] The company describes its models as "audio language models", a single audio-native architecture trained to generate, transcribe and transform speech.[1][2]

Gradium came out of stealth on 2 December 2025 with a $70 million seed round led by FirstMark Capital and Eurazeo.[1][3] On 8 July 2026 it said that an extension of the seed round, with new investors including [Nvidia](https://aiwiki.ai/wiki/nvidia), had brought its total funding to $100 million, and that it was opening an office in the San Francisco Bay Area.[6][7] It has not disclosed a valuation.[8] Its products include the Gradium Translate speech translation models, the on-device TTS model family Phonon (released as open-weight models in October 2026, the smallest under the Apache 2.0 license) and the open-source voice-agent framework Gradbot.[6][9][10]

## Overview

| Field | Value |
|---|---|
| Type | Private company (French SAS)[11] |
| Registered | 15 July 2025, Paris[11] |
| Founding date given by the company | September 2025[1][2] |
| Public launch | 2 December 2025[1] |
| Founders | Neil Zeghidour (CEO), Olivier Teboul, Laurent Mazaré, Alexandre Défossez[1][2] |
| Headquarters | Paris, France; Bay Area office announced July 2026[2][6] |
| Products | Gradium TTS and STT APIs, voice cloning, Voice Design, Gradium Translate (stt-translate, s2s-translate), Phonon on-device TTS, Gradbot, Gradium Studio[6][12][9][10][13] |
| Languages | English, French, German, Spanish, Portuguese[1][12] |
| Funding | $70M seed (December 2025); seed extended to a total of $100M (July 2026)[1][6] |
| Seed leads | FirstMark Capital, Eurazeo[1] |
| Valuation | Not disclosed[8] |
| Website | gradium.ai |

## History

### Background: Google, Meta and Kyutai

Gradium's founders had worked on neural audio codecs and audio [language models](https://aiwiki.ai/wiki/language_model) for several years before starting the company. Zeghidour was the lead author of [SoundStream](https://aiwiki.ai/wiki/soundstream) (2021), a neural audio codec built at Google, and a co-author of [AudioLM](https://aiwiki.ai/wiki/audiolm) (2022), Google's language-modeling approach to audio generation, and of the text-to-music model [MusicLM](https://aiwiki.ai/wiki/musiclm) (2023).[14][15][16] Teboul co-authored LEAF, a learnable audio front end (2021), and AudioLM with Zeghidour.[15][17] At Meta, Défossez was the lead author of the [EnCodec](https://aiwiki.ai/wiki/encodec) codec and the last author of the [MusicGen](https://aiwiki.ai/wiki/musicgen) music-generation paper.[18][19] Kyutai's 2023 launch announcement said Défossez had led research and development for Meta's open-source AudioCraft framework (EnCodec, AudioGen and MusicGen), and that Zeghidour had started and led a generative-audio team at Google DeepMind after three years at Facebook AI Research working on speech recognition.[4]

In November 2023, Défossez, Zeghidour and Mazaré were among the six scientists named at the launch of Kyutai, a nonprofit open-science lab in Paris funded by the Iliad group (Xavier Niel), the CMA CGM group (Rodolphe Saadé) and Schmidt Futures (Eric Schmidt).[4] Mazaré had studied at École Polytechnique, worked as a quantitative strategist at Goldman Sachs, and moved between Jane Street and DeepMind's computer vision group before joining Kyutai.[4] At Kyutai the group co-authored Moshi, a full-duplex spoken dialogue model released in September 2024; Hibiki, a simultaneous speech-to-speech translation model (February 2025); and the Delayed Streams Modeling framework used for streaming speech recognition and synthesis (September 2025).[20][21][22]

### Founding and launch (2025)

On 9 October 2025, Forbes reported that Zeghidour was in talks to raise about $60 million from American and French investors for a new voice AI startup called Gradium, intended to commercialize the voice technology developed at Kyutai. Forbes cited French corporate filings showing that Zeghidour had registered a company in July, and sources who said Kyutai would keep a stake. Zeghidour told Forbes that some details were inaccurate.[23] The French business registry lists GRADIUM, a simplified joint-stock company (SAS) with Zeghidour as president, as created on 15 July 2025 with a registered office in Paris's 1st arrondissement.[11] Gradium itself gives September 2025 as its founding date, and its launch press release described the company as having spent three months in stealth.[1][2]

Gradium launched publicly on 2 December 2025.[1] In an interview with Maddyness, Zeghidour said the spinout was not a divorce from Kyutai: the two organizations shared premises, separated by a corridor, and Kyutai was a Gradium shareholder that would benefit from Gradium's valuation but had no exclusive relationship with it. He said businesses, notably in gaming, had wanted Kyutai's open-source models with wider language coverage, a commercial opportunity that fell outside the nonprofit's research mission.[24] Science|Business reported that Zeghidour would run Gradium full time while Défossez would split his time between Gradium and Kyutai.[5] As of October 2026, Kyutai's own team page still listed Mazaré as head of engineering, Défossez as head of research and Zeghidour as audio research advisor.[25]

At launch, Gradium offered real-time transcription and synthesis in English, French, German, Spanish and Portuguese, sold on plans ranging from developer access to enterprise deployments.[1][2] The company said it had generated its first revenue within weeks of founding, with early customers in gaming, AI agents, customer care, language learning and healthcare; Zeghidour told Maddyness the first customers came after six weeks.[2][24] The French outlet Next reported a starting team of ten people.[26] Speaking at the ai-PULSE conference in Paris on 4 December 2025, Zeghidour said the company's plan was to be "much better than any other solution on the market" on quality, latency, robustness and precision by the end of 2026.[5]

### Growth and seed extension (2026)

Through the first half of 2026 Gradium announced a startup program, partnerships with companies including Wonderful, Acolad and InteractionLabs, a listing on [Amazon Web Services](https://aiwiki.ai/wiki/amazon_web_services), the private beta of Phonon, the open-source Gradbot framework, and an upgraded TTS model.[27][28][29][30][31][32][33][53] On 24 June 2026 it launched Gradium Translate.[12]

On 8 July 2026 Gradium announced that it had "extended its funding to $100 million just seven months after its launch" through an extension of its seed financing, welcoming new investors including Nvidia. The company said the money would fund research, product development, international expansion and a new office in the San Francisco Bay Area.[6] Sifted reported that the extension was about $30 million, bringing the seed round to over $100 million.[7] According to Dealroom News, Gradium declined to share its latest valuation or name the other new investors.[8]

In the second half of 2026 the company replaced its default TTS model twice, added EU and US data residency options, obtained ISO 27001 certification and a SOC 2 Type II attestation, expanded its voice catalog to more than 1,400 voices, and released Phonon as open weights during [NVIDIA GTC Berlin 2026](https://aiwiki.ai/wiki/nvidia_gtc_berlin_2026).[13][9][34][35][36]

## Founders and leadership

| Name | Title at launch (December 2025) | Background |
|---|---|---|
| Neil Zeghidour | Chief Executive Officer | Facebook AI Research, then Google DeepMind, where he started and led a generative-audio team; founding scientist at Kyutai; PhD in machine learning from École Normale Supérieure[1][4] |
| Olivier Teboul | Chief Technology Officer | Google Brain; co-author of LEAF and AudioLM[1][15][17] |
| Laurent Mazaré | Chief Coding Officer | Goldman Sachs, Jane Street, DeepMind; founding scientist and head of engineering at Kyutai; author of the tch-rs Rust bindings for [PyTorch](https://aiwiki.ai/wiki/pytorch)[1][4][25][37] |
| Alexandre Défossez | Chief Science Officer | Meta FAIR research scientist who led AudioCraft (EnCodec, AudioGen, MusicGen); founding scientist at Kyutai[1][4] |

The launch announcement also named Constance Grisoni, formerly of BCG X, as chief growth officer and Eugene Kharitonov, formerly of Meta and Google DeepMind, as founding scientist.[1] Kharitonov is a co-author of AudioLM and of the Delayed Streams Modeling paper.[15][22] Gradium's October 2026 press release for Phonon quoted Mazaré with the title of CTO.[34] Mazaré's GitHub profile describes him as co-founder of both Kyutai and Gradium; he wrote tch-rs and is the largest contributor by commits to Hugging Face's Rust machine learning framework [Candle](https://aiwiki.ai/wiki/candle_hf).[37][38]

## Technology

Gradium describes its models as audio language models, "an audio-native counterpart to large language models" that handles generation, transcription and transformation of speech in one architecture.[1][2] In a May 2026 post the company said its models are based on Delayed Streams Modeling (DSM), a method published by the Kyutai team in September 2025 in which a decoder-only model reads and writes time-aligned text and audio streams with fixed delays between them, so that one formulation covers both streaming speech recognition and streaming TTS.[22][39]

Several Gradium products trace to specific Kyutai papers. The company said stt-translate was "inspired by" Hibiki-Zero, a February 2026 Kyutai paper on simultaneous speech translation that drops the need for word-level aligned training data and uses [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) with [GRPO](https://aiwiki.ai/wiki/grpo) to reduce latency.[12][40] Phonon builds on Kyutai's research on Continuous Audio Language Models (CALM), which generate continuous audio frames instead of discrete codec tokens, the work that led to Kyutai's small CPU-friendly Pocket TTS model.[9][41][47]

## Products

### Text-to-speech and speech-to-text API

Gradium's core product is a streaming API, served over WebSocket and REST, for real-time TTS and STT, with instant voice cloning from a short audio sample and a higher-fidelity "Pro Voice Clone" tier.[32][36][42] Gradium Studio is the web interface for the same models.[13]

- **TTS models.** Gradium shipped an upgraded default TTS model on 10 June 2026 that it said pronounced acronyms, email addresses, phone numbers and codes more accurately.[31] Another new default model followed on 31 August 2026. A further model entered public beta on 22 September 2026, when the company measured time to first audio under 50 ms against 225 ms for the then-production model, and became the default on 30 September.[13][43]
- **Speech recognition.** Gradium's STT model emits turn-completion predictions every 80 ms alongside transcripts, which the company calls semantic voice activity detection: a forecast of whether the user has finished speaking rather than a simple silence detector.[44] Keyword boosting for custom vocabulary was added in July 2026.[13]
- **Voice Design and voice catalog.** Voice Design, launched in September 2026, creates a new voice from a text description.[13] On 7 October 2026 the catalog grew from 387 to 1,421 voices across the five supported languages.[35]
- **Deployment.** Since April 2026 the API has been sold as a managed SaaS subscription on AWS Marketplace, and the TTS model is also available as an [Amazon SageMaker](https://aiwiki.ai/wiki/amazon_sagemaker) image for deployment inside a customer's own cloud account.[32] In September 2026 Gradium added enforced EU or US data residency and a zero-data-retention mode.[36] Its pricing page lists a free tier and four paid plans, plus custom arrangements such as dedicated inference, on-premise deployment, or deployment through [Baseten](https://aiwiki.ai/wiki/baseten) or SageMaker.[45]

Gradium said that as of 13 May 2026 it ranked first on all latency metrics of the TTS benchmark run by the voice-evaluation company Coval.[39]

### Gradium Translate

Gradium Translate, launched on 24 June 2026, consists of two models. stt-translate turns speech in one language into text in another in a single step, without an intermediate transcript, and s2s-translate pairs it with a Gradium TTS voice to produce translated speech, including in a clone of the speaker's own voice. Both cover all 20 directed pairs among English, French, German, Spanish and Portuguese.[12] On a proprietary conversational dataset, Gradium reported higher [BLEU](https://aiwiki.ai/wiki/bleu_bilingual_evaluation_understudy) and MetricX scores than gemini-3.5-live-translate, higher BLEU and comparable MetricX against OpenAI's gpt-realtime-translate, and average latency of 3.0 seconds against 2.9 seconds for Gemini and 3.6 seconds for OpenAI's model. These are the company's own measurements.[12]

### Phonon

Phonon is Gradium's on-device TTS model. It was announced on 3 April 2026 as a private-beta product: Gradium fine-tuned a small model for one customer voice and language and shipped it as a licensed binary that runs on the CPU of Android and iOS devices, with a fixed per-device price instead of per-request API billing.[27]

On 20 October 2026 Gradium released a family of Phonon models with open weights on [Hugging Face](https://aiwiki.ai/wiki/hugging_face), and the next day issued a press release datelined from NVIDIA GTC Berlin.[9][34] The release has three variants:[9][46]

| Model | Parameters | Size (8-bit) | Languages | Built-in voices | License |
|---|---|---|---|---|---|
| Phonon 31m | 31 million (6 transformer layers) | 64 MB | English | 37 | [Apache 2.0](https://aiwiki.ai/wiki/apache_license) |
| Phonon 40m | 40 million (9 transformer layers) | 74.6 MB | One model each for EN, FR, DE, ES, PT | 12 to 37 per language | CC BY-NC 4.0 |
| Phonon 40m multilingual | 40 million | 74.6 MB | EN, FR, DE, ES, PT in one model | 102 | CC BY-NC 4.0 |

Phonon uses a causal [transformer](https://aiwiki.ai/wiki/transformers) backbone conditioned on text embeddings from a [SentencePiece](https://aiwiki.ai/wiki/sentencepiece) tokenizer with a 4,000-token vocabulary, so it reads raw text without phoneme conversion. For each 80 ms frame the backbone produces a latent vector that a [flow-matching](https://aiwiki.ai/wiki/flow_matching) head turns into continuous audio embeddings in one step; a smaller variant of Kyutai's Mimi decoder converts them to 24 kHz audio. All variants are distilled from a larger Phonon teacher model and were trained on [NVIDIA H100](https://aiwiki.ai/wiki/nvidia_h100) GPUs.[9] The inference code, released under Apache 2.0 in the gradium-ai/xn-ptts repository, runs on CPUs and through CUDA, Metal, Vulkan, WebGPU and Core ML backends, with bindings for Rust, Python, Swift and [WebAssembly](https://aiwiki.ai/wiki/webassembly).[9][10]

Gradium reported that Phonon 31m reached 1.70% [word error rate](https://aiwiki.ai/wiki/word_error_rate) on the English part of a public test set of 500 hard sentences per language (phone numbers, email addresses, codes and similar cases), against 3.50% for Kokoro, and that Phonon 31m started speaking in about 7 ms on a MacBook Air M5 CPU against 517 ms for Kokoro, which does not stream and so must synthesize the whole sentence first. Its blind listening comparison used 1,189 raters and 17,270 votes.[9] Launch partners named in the press release were Hugging Face, Arm, Qualcomm, Appning by FORVIA, LM Studio, [Ollama](https://aiwiki.ai/wiki/ollama) and MLX Audio, and FORVIA was described as a design partner using Phonon in automotive research.[9][34] AlphaSignal noted that the benchmark results are "company-published rather than an independent evaluation".[47]

### Gradbot

Gradbot is an open-source framework (dual-licensed under MIT or Apache 2.0), released on 15 April 2026, for prototyping voice agents. A multiplexing core written in Rust coordinates the speech-to-text, language-model and text-to-speech streams and handles turn-taking, interruptions, silences and asynchronous tool calls, so that a working agent takes about 50 lines of code. Gradium positions it for prototypes and recommends orchestrators such as [LiveKit](https://aiwiki.ai/wiki/livekit_agents) and [Pipecat](https://aiwiki.ai/wiki/pipecat) for production.[28][10]

## Customers and partnerships

| Date | Partner | Arrangement |
|---|---|---|
| January 2026 | Wonderful | Gradium models power real-time voice agents on Wonderful's platform[29] |
| January 2026 | Invincible Voice (Kyutai) | Gradium's API powers an open-source assistive voice system, led by Kyutai, for people with ALS or other speech loss[48] |
| February 2026 | Acolad | Partnership on AI interpreting for enterprise and public-sector use[30] |
| February 2026 | InteractionLabs | Gradium voices, STT and TTS power the Ongo home robot[33] |
| April 2026 | AWS | SaaS listing on AWS Marketplace and a SageMaker model image[32] |
| March 2026 (announced July 2026) | RMC BFM | TTS with Pro Voice Clones of RMC BFM journalists for RMC BFM Drive, an AI-generated personalized radio app on Renault's openR link system[42] |
| August 2026 | AudioStack | Gradium voices added to AudioStack's audio production platform[49] |
| September 2026 | LiveKit | Gradium TTS available on LiveKit Inference[50] |
| October 2026 | Eden AI | TTS, STT and speech translation available through Eden AI's gateway, pinned to the EU region[51] |
| October 2026 | FORVIA / Appning by FORVIA | Design partners for Phonon in automotive research[9] |
| October 2026 | Yampa | Gradium TTS and STT run natively in Yampa's customer-care agent platform[52] |

Gradium's startup program, launched in February 2026, gives eligible seed or Series A startups with fewer than 25 employees six months of its M plan.[53]

## Funding

| Date | Round | Amount | Investors | Valuation |
|---|---|---|---|---|
| October 2025 | Reported talks | About $60M sought (Forbes) | Not named | Not reported[23] |
| 2 December 2025 | Seed | $70M (about €60M as reported in French press) | Led by FirstMark Capital and Eurazeo; with DST Global Partners, Eric Schmidt, Xavier Niel, Rodolphe Saadé, Korelya Capital, Amplify Partners, Liquid2, Drysdale Ventures; angels including [Yann LeCun](https://aiwiki.ai/wiki/yann_lecun), Olivier Pomel, Ilkka Paananen, Thomas Wolf, Guillermo Rauch and Mehdi Ghissassi | Not disclosed[1][24][26] |
| 8 July 2026 | Seed extension | About $30M per Sifted; total funding $100M per Gradium | New investors including Nvidia; others not named | Not disclosed[7][6][8] |

TechCrunch and Sifted reported the December round as $70 million; Maddyness reported it as €60 million, and Next gave both figures ($70 million, about €60 million).[3][24][26][54] The company has not published a valuation for either the seed round or the extension.[8]

## Security and compliance

Gradium's API runs in EU and US regions. With data residency enabled, a session is processed only in the chosen region and is refused rather than rerouted if it reaches the wrong one; account data and custom voices are stored in an EU control plane for all customers.[36] Gradium obtained ISO 27001 certification in September 2026 and a SOC 2 Type II attestation in October 2026.[13] Its website says it is going through France's HDS certification for hosting personal health data.[55]

## Reception and competition

TechCrunch's launch coverage noted that Gradium was entering a crowded field that included voice and multimodal models from [OpenAI](https://aiwiki.ai/wiki/openai), [Anthropic](https://aiwiki.ai/wiki/anthropic), Meta and [Mistral AI](https://aiwiki.ai/wiki/mistral_ai), well-funded startups such as [ElevenLabs](https://aiwiki.ai/wiki/elevenlabs), and hundreds of open speech models on Hugging Face.[3] Forbes wrote that Gradium would go head to head with London-based ElevenLabs.[23] According to Dealroom News, Zeghidour said in July 2026 that "less than a dozen players" could train voice models at scale and that Gradium published new models "almost every month", against six-to-twelve-month cycles at rivals.[8] Gradium's own benchmark posts compare its TTS against [Cartesia](https://aiwiki.ai/wiki/cartesia), [Inworld AI](https://aiwiki.ai/wiki/inworld_ai) and ElevenLabs models.[39][31]

Debarghya Das's "Ultimate Neolab List" (October 2026) includes Gradium and rates it "Borderline", citing its research pedigree and own models against its position as an API product company with modest funding and compute.[56]

## See also

- [List of neolabs](https://aiwiki.ai/wiki/list_of_neolabs)
- [Neolab](https://aiwiki.ai/wiki/neolab)
- [Kyutai](https://aiwiki.ai/wiki/kyutai_labs)
- [Moshi](https://aiwiki.ai/wiki/moshi)
- [Text-to-Speech](https://aiwiki.ai/wiki/text_to_speech_ai)
- [Text-to-Speech Models](https://aiwiki.ai/wiki/text-to-speech_models)
- [AI Voice Agent](https://aiwiki.ai/wiki/ai_voice_agent)
- [ElevenLabs](https://aiwiki.ai/wiki/elevenlabs)
- [Cartesia](https://aiwiki.ai/wiki/cartesia)

## References

1. Gradium. "Gradium: Solving voice." Gradium blog, December 2, 2025. https://gradium.ai/blog/gradium
2. Gradium. "Gradium Launches to Redefine Voice AI" (press release, English). December 2, 2025. https://gradium.ai/Gradium%20-%20press%20release%20EN%20final.pdf
3. Julie Bort. "Paris-based AI voice startup Gradium nabs $70M seed." TechCrunch, December 2, 2025. https://techcrunch.com/2025/12/02/paris-based-ai-voice-startup-gradium-nabs-70m-seed/
4. Kyutai. "Launch of Kyutai: Europe's first independent research lab dedicated to AI open science" (press release). November 17, 2023. https://kyutai.org/CP_Kyutai_AI_EN.pdf
5. Martin Greenacre. "How a non-profit research lab produced France's hottest new AI start-up." Science|Business, December 9, 2025 (archived February 1, 2026). https://web.archive.org/web/20260201142522/https://sciencebusiness.net/news/r-d-funding/ai/how-non-profit-research-lab-produced-frances-hottest-new-ai-start
6. Gradium. "Gradium Extends Funding to $100 Million and Expands to Silicon Valley." Gradium blog, July 8, 2026. https://gradium.ai/blog/gradium-100-million-funding-nvidia
7. Daphné Leprince-Ringuet. "Nvidia backs voice AI startup Gradium, bringing seed round to over $100m." Sifted, July 8, 2026. https://sifted.eu/articles/gradium-nvidia-30m-extension-seed
8. Dealroom News. "Gradium adds $30M from Nvidia, pushing seed round past $100M." July 8, 2026. https://dealroom.co/news/137701-gradium-adds-30m-from-nvidia-pushing-seed-round-past-100m/
9. Gradium. "Open-Sourcing Phonon: A Suite of Frontier On-Device TTS Models." Gradium blog, October 20, 2026. https://gradium.ai/blog/release-phonon-open-weight
10. Gradium. "gradium-ai" organization repositories (gradbot, xn, xn-ptts, gradium-py, gradium-rs). GitHub. Accessed October 27, 2026. https://github.com/gradium-ai
11. French government business registry (Recherche d'entreprises API). "GRADIUM, SIREN 989284955." Accessed October 27, 2026. https://recherche-entreprises.api.gouv.fr/search?q=gradium
12. Gradium. "Launching Gradium Translate: the best accuracy-latency tradeoff against gemini-3.5-live-translate and gpt-realtime-translate." Gradium blog, June 24, 2026. https://gradium.ai/blog/launching-gradium-translate
13. Gradium. "Release Notes." Gradium API documentation. Accessed October 27, 2026. https://docs.gradium.ai/guides/release-notes
14. Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, Marco Tagliasacchi. "SoundStream: An End-to-End Neural Audio Codec." arXiv:2107.03312, July 7, 2021. https://arxiv.org/abs/2107.03312
15. Zalán Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, Neil Zeghidour. "AudioLM: a Language Modeling Approach to Audio Generation." arXiv:2209.03143, September 7, 2022. https://arxiv.org/abs/2209.03143
16. Andrea Agostinelli et al. "MusicLM: Generating Music From Text." arXiv:2301.11325, January 26, 2023. https://arxiv.org/abs/2301.11325
17. Neil Zeghidour, Olivier Teboul, Félix de Chaumont Quitry, Marco Tagliasacchi. "LEAF: A Learnable Frontend for Audio Classification." arXiv:2101.08596, January 21, 2021. https://arxiv.org/abs/2101.08596
18. Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi. "High Fidelity Neural Audio Compression." arXiv:2210.13438, October 24, 2022. https://arxiv.org/abs/2210.13438
19. Jade Copet, Felix Kreuk, Itai Gat, Tal Remez, David Kant, Gabriel Synnaeve, Yossi Adi, Alexandre Défossez. "Simple and Controllable Music Generation." arXiv:2306.05284, June 8, 2023. https://arxiv.org/abs/2306.05284
20. Alexandre Défossez, Laurent Mazaré, Manu Orsini, Amélie Royer, Patrick Pérez, Hervé Jégou, Edouard Grave, Neil Zeghidour. "Moshi: a speech-text foundation model for real-time dialogue." arXiv:2410.00037, September 17, 2024. https://arxiv.org/abs/2410.00037
21. Tom Labiausse, Laurent Mazaré, Edouard Grave, Patrick Pérez, Alexandre Défossez, Neil Zeghidour. "High-Fidelity Simultaneous Speech-To-Speech Translation." arXiv:2502.03382, February 5, 2025. https://arxiv.org/abs/2502.03382
22. Neil Zeghidour, Eugene Kharitonov, Manu Orsini, Václav Volhejn, Gabriel de Marmiesse, Edouard Grave, Patrick Pérez, Laurent Mazaré, Alexandre Défossez. "Streaming Sequence-to-Sequence Learning with Delayed Streams Modeling." arXiv:2509.08753, September 10, 2025. https://arxiv.org/abs/2509.08753
23. Iain Martin. "Former Meta And Google Researcher In Talks To Raise $60 Million For Voice AI Startup." Forbes, October 9, 2025. https://www.forbes.com/sites/iainmartin/2025/10/09/former-meta-and-google-researcher-in-talks-to-raise-60-million-for-voice-ai-startup/
24. Maxence Fabrion. "Gradium, premier spin-off de Kyutai, lève 60 millions d'euros pour devenir une référence mondiale de l'IA vocale." Maddyness, December 2, 2025. https://www.maddyness.com/2025/12/02/gradium-premier-spin-off-de-kyutai-leve-60-millions-deuros-pour-devenir-une-reference-mondiale-de-lia-vocale/
25. Kyutai. "Team." Accessed October 27, 2026. https://kyutai.org/team
26. Alexandre Laurent. "Issue de Kyutai, la startup Gradium lève 60 millions d'euros pour ses IA vocales." Next, December 2, 2025. https://next.ink/brief-article/issue-de-kyutai-la-startup-gradium-leve-60-millions-deuros-pour-ses-ia-vocales/
27. Gradium. "Gradium Phonon: On-Device TTS for Consumer Apps, NPCs, and Offline Products." Gradium blog, April 3, 2026. https://gradium.ai/blog/gradium-phonon
28. Gradium. "Gradbot: Vibe code voice agents in 50 lines of code." Gradium blog, April 15, 2026. https://gradium.ai/blog/gradbot
29. Gradium. "Powering Wonderful's Voice Agents." Gradium blog, January 15, 2026. https://gradium.ai/blog/wonderful
30. Gradium. "Acolad and Gradium Partner to Advance Enterprise-Ready AI Interpreting." Gradium blog, February 5, 2026. https://gradium.ai/blog/acolad-gradium-partnership
31. Gradium. "Gradium TTS, upgraded: more accurate Text-To-Speech." Gradium blog, June 10, 2026. https://gradium.ai/blog/gradium-tts-upgrade
32. Gradium. "Gradium Voice Launches on AWS as a SaaS Subscription and a SageMaker Model Image." Gradium blog, April 29, 2026. https://gradium.ai/blog/gradium-aws-launch
33. Gradium. "InteractionLabs (Ongo) and Gradium Partner to Redefine Human-Robot Interaction." Gradium blog, February 17, 2026. https://gradium.ai/blog/interactionlabs-gradium-partnership
34. Gradium. "Gradium releases Phonon, an open-source text-to-speech model that runs on any device." Gradium blog (press release), October 21, 2026. https://gradium.ai/blog/gradium-releases-phonon
35. Gradium. "1,000+ new voices, our most natural set yet." Gradium blog, October 7, 2026. https://gradium.ai/blog/more-than-1000-voices
36. Gradium. "Build voice agents with EU and US residency and Zero Data Retention." Gradium blog, September 2, 2026. https://gradium.ai/blog/eu-data-residency-zero-data-retention
37. Laurent Mazaré. GitHub profile and repositories (LaurentMazare/tch-rs; contributions to huggingface/candle). GitHub. Accessed October 27, 2026. https://github.com/LaurentMazare
38. GitHub. "huggingface/candle" contributors list (GitHub REST API). Accessed October 27, 2026. https://api.github.com/repos/huggingface/candle/contributors
39. Gradium. "Gradium #1 on Coval TTS Benchmarks." Gradium blog, May 13, 2026. https://gradium.ai/blog/coval-tts-benchmarks-may-2026
40. Tom Labiausse, Romain Fabre, Yannick Estève, Alexandre Défossez, Neil Zeghidour. "Simultaneous Speech-to-Speech Translation Without Aligned Data." arXiv:2602.11072, February 11, 2026. https://arxiv.org/abs/2602.11072
41. Simon Rouard, Manu Orsini, Axel Roebel, Neil Zeghidour, Alexandre Défossez. "Continuous Audio Language Models." arXiv:2509.06926, September 8, 2025. https://arxiv.org/abs/2509.06926
42. Gradium. "Gradium Powers RMC BFM Drive: AI-Generated Personalized Radio in Renault Vehicles." Gradium blog, July 7, 2026. https://gradium.ai/blog/rmc-bfm-drive-gradium
43. Gradium. "Why we built the fastest robust TTS model." Gradium blog, October 1, 2026. https://gradium.ai/blog/fastest-robust-tts-model
44. Gradium. "Semantic VAD: turn detection that uses meaning, not silence." Gradium blog, June 2, 2026. https://gradium.ai/blog/semantic-vad
45. Gradium. "Pricing." Accessed October 27, 2026. https://gradium.ai/pricing
46. Gradium. "gradium/phonon-31m-en" and "gradium/phonon-40m-multilingual" model cards. Hugging Face. Accessed October 27, 2026. https://huggingface.co/gradium/phonon-31m-en
47. AlphaSignal. "Gradium's Phonon Runs Text-to-Speech in Browsers at Under 75 MB." October 20, 2026. https://alphasignal.ai/news/gradium-s-phonon-runs-text-to-speech-in-browsers-at-under-75-mb
48. Gradium. "Invincible Voice: How Gradium's Real-Time Voice AI Helps ALS Patients Speak Again." Gradium blog, January 28, 2026. https://gradium.ai/blog/invincible-voice
49. Gradium. "Gradium voices are now live on AudioStack." Gradium blog, August 19, 2026. https://gradium.ai/blog/audiostack-gradium-partnership
50. Gradium. "Gradium now available for free on LiveKit Inference." Gradium blog, September 9, 2026. https://gradium.ai/blog/livekit-gradium-partnership
51. Gradium. "Gradium is now available on Eden AI." Gradium blog, October 19, 2026. https://gradium.ai/blog/eden-ai-gradium-partnership
52. Gradium. "Gradium and Yampa partner on voice AI agents for customer experience." Gradium blog, October 26, 2026. https://gradium.ai/blog/yampa-gradium-partnership
53. Gradium. "Building Voice Agents From the Ground Up: The Gradium Startup Program." Gradium blog, February 9, 2026. https://gradium.ai/blog/gradium-startup-program
54. Daphné Leprince-Ringuet. "Ex-Google CEO Eric Schmidt backs voice AI startup Gradium in $70m seed round." Sifted, December 2, 2025. https://sifted.eu/articles/gradium-70m-seed-voice-ai
55. Gradium. "Health Data Hosting (HDS)." Accessed October 27, 2026. https://gradium.ai/hds
56. Debarghya Das. "The Ultimate Neolab List." Accessed October 27, 2026. https://debarghyadas.com/neolabs/