To help you find voice models with real-time speech-to-speech capabilities similar to GPT-4o, could you please clarify the following:

1. Are you looking for open-source models, commercial APIs, or both?
2. What level of latency is acceptable for your application (in milliseconds or qualitative terms like "near real-time")?
3. What are your deployment preferences—on-premise, edge device, or cloud?
4. Any specific languages or accents you need supported?
5. Are there other technical constraints (e.g., hardware limitations, privacy requirements)?

Once I have this, I’ll begin the search for voice models that can support your use case effectively.

Got it. I’ll look for voice models that offer real-time speech-to-speech capabilities similar to GPT-4o, with a focus on minimizing latency beyond traditional VTT-LLM-TTS pipelines. I’ll include both open-source and commercial options, and highlight their latency, architecture, deployment models, and language support.

I’ll update you shortly with the most promising models and platforms for building low-latency voice applications.

Real-Time Speech-to-Speech Models (Integrated S2S)

Modern speech AI is moving toward integrated speech-to-speech (S2S) models that can directly convert incoming speech to a voice response with minimal delay. These models avoid the traditional cascade of separate voice-to-text (ASR), text-based LLM, and text-to-speech components, which often introduce latency. Below, we explore notable models and platforms with real-time S2S capabilities similar to OpenAI’s GPT-4o (an “omni-modal” GPT-4). We compare their architecture, latency, language support, deployment options, and more.

OpenAI GPT-4o (Realtime Multi-Modal GPT-4)

Architecture: GPT-4o is a unified multimodal model that handles speech input and generates spoken output directly, without an external ASR or TTS module in the loop (OpenAI Realtime API: The Missing Manual) (OpenAI Realtime API: The Missing Manual) . It processes audio, text, and images and can output text or synthesized speech in one end-to-end system (Hello GPT-4o | OpenAI) . This greatly simplifies voice pipelines.

Latency: Extremely low. GPT-4o can begin responding to voice input in as little as 232 ms (average ~320 ms) (Hello GPT-4o | OpenAI) – comparable to human conversational response times. This is a massive improvement over earlier GPT-4 voice pipelines that took 1–3 seconds total (OpenAI Realtime API: The Missing Manual) . The end-to-end model streams out its answer, achieving near real-time interaction.

Languages & Accents: It matches GPT-4 Turbo’s prowess in English and even improves on GPT-4 in many non-English languages (Hello GPT-4o | OpenAI) . In practice, GPT-4o can understand and respond in multiple languages (it was designed for broad language understanding). It uses high-quality neural voices, making outputs sound natural. (OpenAI’s voice tech can speak with different styles or accents, though details on accent variety are limited in documentation.)

Deployment: Available via OpenAI’s cloud (Realtime API). Developers can stream audio into GPT-4o and get streaming audio out (OpenAI Realtime API: The Missing Manual) . There’s no on-prem or offline option – the model runs on OpenAI’s servers. The API manages the streaming, turn-taking, and state for conversations (OpenAI Realtime API: The Missing Manual) , so developers don’t need to orchestrate ASR/LLM/TTS manually.

SDK/API: Yes – OpenAI provides a Realtime API (preview) for GPT-4o with WebSocket or similar streaming support (OpenAI Realtime API: The Missing Manual) . This makes integration straightforward.

Use Cases: Ideal for voice assistants, interactive voice agents, or AI voice chat in applications where low latency is critical. For example, GPT-4o powers the voice mode in ChatGPT, enabling fluid back-and-forth dialogue. Demos have shown it handling phone calls and voice conversations with human-like response speed and reasoning.

Amazon Nova Sonic (Bedrock Voice Foundation Model)

Architecture: Amazon’s Nova Sonic is a unified speech model that merges speech recognition, understanding, and generation into a single end-to-end model (Introducing Amazon Nova Sonic: Human-like voice conversations for generative AI applications | AWS News Blog) . It is part of AWS Bedrock’s foundation models. Nova Sonic simultaneously transcribes incoming speech and generates an audio reply using one model, preserving nuances like tone and speaking style (Introducing Amazon Nova Sonic: Human-like voice conversations for generative AI applications | AWS News Blog) . This integrated approach avoids the complexity of chaining Alexa’s old ASR/NLU/TTS pipeline.

Latency: Designed for real-time performance. Nova Sonic supports bidirectional streaming, so it can start speaking while you’re still talking, enabling fluid, low-latency conversations (Introducing Amazon Nova Sonic: Human-like voice conversations for generative AI applications | AWS News Blog) (Introducing Amazon Nova Sonic: Human-like voice conversations for generative AI applications | AWS News Blog) . While exact figures aren’t public, it’s optimized for <500 ms end-to-end latency. Amazon touts “low latency and industry-leading price-performance” for the model (Introducing Amazon Nova Sonic: Human-like voice conversations for generative AI applications | AWS News Blog) . In practice, conversations feel natural with immediate turn-taking (a demo showed smooth interruption handling and <1s response times).

Languages & Accents: At launch (Apr 2025), it supports English (with robust understanding of American and British English) (Introducing Amazon Nova Sonic: Human-like voice conversations for generative AI applications | AWS News Blog) . It generates highly expressive speech in a natural US or UK accent, adjusting its delivery based on the user’s prosody. Amazon plans to add more languages soon (Introducing Amazon Nova Sonic: Human-like voice conversations for generative AI applications | AWS News Blog) . The model captures vocal cues from input – for example, it can modulate response style (pace, timbre) to match or respond appropriately to the user’s tone (Introducing Amazon Nova Sonic: Human-like voice conversations for generative AI applications | AWS News Blog) . Built-in content moderation is included for safe deployment (Introducing Amazon Nova Sonic: Human-like voice conversations for generative AI applications | AWS News Blog) .

Deployment: Cloud only, via Amazon Bedrock. Nova Sonic can be invoked through Bedrock’s new streaming API (InvokeModelWithBidirectionalStream) which lets developers send audio in and get audio out in real time (Introducing Amazon Nova Sonic: Human-like voice conversations for generative AI applications | AWS News Blog) . Being a managed AWS service, it won’t run locally, but AWS’s global infrastructure ensures low-latency access in many regions.

SDK/API: Yes – available through AWS SDK/Bedrock API. Developers enable the model in Bedrock and integrate streaming audio I/O calls. Nova Sonic supports tool use and function calling via Bedrock (e.g., it can fetch data or use other APIs mid-dialogue) (Introducing Amazon Nova Sonic: Human-like voice conversations for generative AI applications | AWS News Blog) (Introducing Amazon Nova Sonic: Human-like voice conversations for generative AI applications | AWS News Blog) , which is useful for building agentive voice assistants (e.g. querying databases during a call).

Use Cases: Geared toward enterprise voice agents and conversational AI. For example, it’s demoed in a contact center scenario handling a customer support call, with real-time sentiment analysis and the model dynamically adjusting its empathetic tone (Introducing Amazon Nova Sonic: Human-like voice conversations for generative AI applications | AWS News Blog) (Introducing Amazon Nova Sonic: Human-like voice conversations for generative AI applications | AWS News Blog) . It’s suitable for smart assistants, voice bots in customer service, interactive education apps, and any voice-driven application needing natural dialogue with low latency.

Microsoft Azure Voice Live API (Preview)

Architecture: Azure’s Voice Live API provides an all-in-one speech-to-speech service that unifies speech recognition, an LLM-based dialog engine, and text-to-speech in a single interface (Voice Live API overview - Azure AI services | Microsoft Learn) . Instead of developers manually connecting Azure Speech-to-Text, Azure OpenAI, and Text-to-Speech, the Voice Live API orchestrates it under the hood. This consolidation reduces complexity and latency (Voice Live API overview - Azure AI services | Microsoft Learn) .

Latency: A key goal is ultra-low latency interactive speech. By streaming audio in and out through one pipeline, Azure minimizes end-user perceived delay (Voice Live API overview - Azure AI services | Microsoft Learn) . Although precise latency metrics aren’t given, the design is for near-instant responses (likely on the order of a few hundred milliseconds, enough for natural turn-taking). Microsoft specifically notes this solution aims to deliver low latency, real-time voice interactions without the typical multi-component delays (Voice Live API overview - Azure AI services | Microsoft Learn) .

Languages & Accents: The preview doesn’t list supported languages publicly, but it builds on Azure’s Speech services which cover 100+ languages for recognition and dozens of neural voices for output. We can expect English support by default and likely others (Azure’s TTS offers many accents—US, UK, etc., and languages from Arabic to Japanese). The unified model can presumably handle any language that the underlying models support, but initial focus may be on English voice agents. As it’s a managed service, developers can probably choose output voice personas (male/female, different locales) from Azure’s catalogue for the TTS part.

Deployment: Cloud (Azure). Currently in public preview (as of mid-2025) (Voice Live API overview - Azure AI services | Microsoft Learn) , the Voice Live API is accessed via Azure’s endpoint. It streams audio over secure channels and runs in Microsoft’s cloud. There’s no offline deployment, but Azure’s global data centers and edge optimizations should ensure low latency to client devices.

SDK/API: Yes – Azure provides REST and WebSocket streaming APIs for the Voice Live service. Developers simply send audio and receive incremental transcripts and audio output. The API handles turn-end detection, interrupt handling, etc., making it easier to build voice bots (Voice Live API overview - Azure AI services | Microsoft Learn) . (Microsoft’s documentation emphasizes not needing to manage multiple components – just call the unified API (Voice Live API overview - Azure AI services | Microsoft Learn) .)

Use Cases: Contact center bots, IVR systems, voice assistants, and kiosk/robot interfaces. Microsoft highlights customer support lines and self-service voice bots as prime scenarios (Voice Live API overview - Azure AI services | Microsoft Learn) . Essentially, anywhere you need a conversational agent that listens and talks back instantly – e.g. virtual receptionists, in-car assistants – this API can be applied. It’s also attractive for developers already in the Azure ecosystem who want a fast voice-enabled chatbot without stitching together services.

Meta SeamlessM4T (Multilingual Speech Translator)

Architecture: SeamlessM4T is an open-source end-to-end S2S model from Meta that handles speech-to-text and speech-to-speech translation in a unified architecture (SeamlessM4T-v2) . It was trained on a massively multilingual dataset to allow direct speech input to either translated text or spoken output. Unlike pipelines that ASR then translate then TTS, SeamlessM4T’s single-model approach preserves vocal characteristics and context across translation. It has specialized variants: SeamlessStreaming (for low-latency simultaneous translation) and SeamlessExpressive (which preserves speaker style/prosody in the output) (SeamlessM4T-v2) .

Latency: The model is designed for streaming, enabling near real-time translation without waiting for full sentences. In fact, it introduces an attention-based simultaneous translation mechanism so that it can start generating target speech even as it continues to listen to the source speech ([2312.05187] Seamless: Multilingual Expressive and Streaming Speech Translation) . This yields low latency – potentially on the order of a few hundred milliseconds for partial outputs. (It’s the first of its kind to do multi-language speech simultaneous translation ([2312.05187] Seamless: Multilingual Expressive and Streaming Speech Translation) .) Actual latency depends on hardware, but Meta’s research goal was sub-500 ms delays for interactive use.

Languages & Accents: Extremely broad. SeamlessM4T v2 was trained on data covering 76+ languages (SeamlessM4T-v2) (spanning widely spoken tongues and low-resource languages). It can therefore understand and translate many language pairs. The output speech is generated via a multilingual acoustic model (vocoder). While you can’t pick a specific accent or voice, the model can preserve the input speaker’s vocal traits to some degree when translating (the Expressive mode attempts to maintain voice style and emotion (SeamlessM4T-v2) ). This is valuable for use cases like translating a speaker’s speech while keeping their tone. It supports a variety of accents natively in the sense that it was trained on diverse speakers – so it’s robust to different accents in input.

Deployment: Open-source (model weights available). Developers can run SeamlessM4T on-premise or on edge devices with sufficient AI hardware (it’s a large Transformer model, so typically it would run on a GPU). Meta has released models (e.g. facebook/seamless-m4t-v2-large on HuggingFace) that can be used under a research license (SeamlessM4T-v2) . There are also demos on platforms like HuggingFace Spaces and Replicate. Deployment flexibility is high – you could host it in your own cloud or locally, though real-time performance will depend on having a powerful machine or optimizing the model (possibly using GPU inference servers).

SDK/API: No official service from Meta, but the model can be used via libraries. For example, Hugging Face’s Transformers provides an interface to SeamlessM4T (SeamlessM4T-v2) . With that, one can feed audio in and get text or audio out in a few lines of code. Some third parties might offer APIs (for instance, an open API on Replicate or others wrapping the model). However, out-of-the-box it’s more of a developer-managed model rather than a ready SaaS API.

Use Cases: Excellent for real-time translation devices, multilingual assistants, and cross-lingual communication. For instance, it could power an interpreter app that lets people speaking different languages converse fluidly, or a live translation feature in video calls. Its ability to keep the speaker’s voice characteristics is useful for translating videos or speeches while preserving the original speaker’s identity. It’s also a foundation for research into voice-based agents that can operate in many languages.

LLaMA-Omni (Open-Source GPT-4o Alternative)

Architecture: LLaMA-Omni is an open-source end-to-end speech interaction LLM built on Llama-3.1-8B (an 8-billion-parameter model) (GitHub - ictnlp/LLaMA-Omni: LLaMA-Omni is a low-latency and high-quality end-to-end speech interaction model built upon Llama-3.1-8B-Instruct, aiming to achieve speech capabilities at the GPT-4o level.) . It extends a text LLM to handle audio input and output, aiming to achieve GPT-4o-level voice capabilities in a much smaller package (GitHub - ictnlp/LLaMA-Omni: LLaMA-Omni is a low-latency and high-quality end-to-end speech interaction model built upon Llama-3.1-8B-Instruct, aiming to achieve speech capabilities at the GPT-4o level.) . The model was post-trained to accept spoken prompts and simultaneously generate text and speech responses (GitHub - ictnlp/LLaMA-Omni: LLaMA-Omni is a low-latency and high-quality end-to-end speech interaction model built upon Llama-3.1-8B-Instruct, aiming to achieve speech capabilities at the GPT-4o level.) . This means one model produces both the transcribed answer and the spoken audio output in one go – similar in spirit to GPT-4o.

Latency: Ultra-fast. Thanks to its compact size and optimizations, LLaMA-Omni demonstrates response latency as low as ~226 ms end-to-end (GitHub - ictnlp/LLaMA-Omni: LLaMA-Omni is a low-latency and high-quality end-to-end speech interaction model built upon Llama-3.1-8B-Instruct, aiming to achieve speech capabilities at the GPT-4o level.) , nearly matching GPT-4o’s responsiveness. It supports streaming generation of speech, enabling the system to start talking almost immediately after the user stops speaking. The low latency makes it feasible for real-time conversations on consumer hardware (with an 8B model, a decent GPU can handle it).

Languages & Accents: Primarily English for now (the project targets English voice conversations, though the name “LLaMA” suggests the base model had some multilingual capability). The focus was on seamless dialogue, so it likely understands English input and responds with a single default synthesized voice. That voice is a neural voice that the model learned during training (quality is high for an 8B model, though not as polished as big commercial TTS). There isn’t mention of multi-language support; LLaMA-Omni’s goal was to replicate ChatGPT’s voice mode, so it emphasizes English conversational use. Being open-source, it could be extended to other languages via fine-tuning on multilingual data in the future.

Deployment: Self-hosted – available on GitHub under Apache-2.0 license (GitHub - ictnlp/LLaMA-Omni: LLaMA-Omni is a low-latency and high-quality end-to-end speech interaction model built upon Llama-3.1-8B-Instruct, aiming to achieve speech capabilities at the GPT-4o level.) . Anyone can download the model and run it on local hardware. With 8B parameters, it can run on a single high-end GPU or even CPU (with slower performance). This gives tremendous deployment flexibility: developers can embed it in edge devices or offline applications where cloud access is undesired. The trade-off is that it’s not as powerful in general reasoning as GPT-4, but it’s lightweight enough for personal use cases.

SDK/API: No official API, but since it’s open-source, developers can integrate it using the provided code. The repository includes examples and even a Cog/Replicate config, indicating it can be run via container or on Replicate’s API easily. Essentially, one would use the model in a Python environment (via HuggingFace transformers or the repo’s code) to pass audio input and get generated speech output. Enthusiasts have likely containerized it for easy use in voice assistant projects.

Use Cases: Offline or private voice assistants, IoT voice interfaces, and research. LLaMA-Omni is suitable for hobbyists or companies that need a voice AI but must run fully on-prem (for privacy or latency reasons). It’s also a testbed for academic research into end-to-end spoken dialogue. Because it produces both text and audio, applications can use the text output for logging or further NLP while the user hears the spoken answer. Its low resource requirement makes it one of the few edge-capable S2S assistants.

Google Translatotron 2 (Speech Translator, Research)

Architecture: Translatotron 2 is Google’s research project for direct speech-to-speech translation (Translatotron 2: Robust direct speech-to-speech translation) . Unlike conventional speech translators that do speech→text→text→speech, Translatotron is a single sequence-to-sequence model that takes input speech and generates translated speech in the target language. Internally it uses a speech encoder, a phoneme decoder, and an acoustic synthesizer, all connected by one attention mechanism (Translatotron 2: Robust direct speech-to-speech translation) . This end-to-end design means it effectively “thinks” in speech units. A notable feature is the ability to retain the source speaker’s voice in the output – so if a person speaks Spanish, the English output sounds like the same person speaking English (Translatotron 2: Robust direct speech-to-speech translation) (they achieved this by predicting vocoder features that preserve voice characteristics, while mitigating misuse by disallowing arbitrary voice impersonation).

Latency: Designed for quality rather than ultra-streaming, but still faster than a multi-stage pipeline. Because it doesn’t have to generate an intermediate text, it can start generating the translation once enough audio is heard to determine the translation. In research settings, they likely processed whole sentences; however, the model could be run in a streaming fashion with some modifications. Latency numbers aren’t published, but being end-to-end it avoids cumulative delays. We can infer it’s capable of near real-time translation on powerful hardware (especially since it was aimed at practical deployment by improving robustness (Translatotron 2: Robust direct speech-to-speech translation) ).

Languages & Accents: In research experiments, Translatotron 2 was tested on Spanish-to-English (and possibly other language pairs) – the original Translatotron was demonstrated on Spanish↔English. It could in principle be trained for many language pairs. The key aspect is accent/voice preservation: it keeps the speaker’s timbre and speaking style in the translated speech (Translatotron 2: Robust direct speech-to-speech translation) . This means the output English will carry an accent influenced by the Spanish speaker’s voice, yielding a more authentic communication (though one might argue it introduces a slight accent in the output, it preserves identity). The model is not known to handle multiple output voice choices – it specifically tries to use the source speaker’s voice as the output voice, which is unique among S2S models.

Deployment: Currently research-only. There is no public API or product for Translatotron 2. Google has not (as of the latest information) deployed it in Google Translate or devices. However, the concept could be integrated into future translators. Some parts of the model (like the idea of units or the vocoder) have influenced other projects. Developers keen to experiment would have to read the research paper and possibly find unofficial implementations. It’s not readily available like SeamlessM4T.

SDK/API: None publicly. It’s mentioned here as a state-of-the-art example of integrated S2S. If Google integrates it into a product, it would likely be via their Cloud Speech Translation API or a future Pixel phone feature. As of now, it remains an advanced prototype (with maybe internal use at Google).

Use Cases: The main use is speech translators – imagine a device or app that instantly translates conversations, with each person’s voice translated into the other’s language. It’s also valuable for dubbing videos or live speeches while keeping the original speaker’s voice (making the experience more personal). By eliminating text, it could eventually power faster, more natural translation services (no awkward pauses for intermediate transcription).

NVIDIA Riva (Optimized Speech AI Pipeline)

Architecture: NVIDIA Riva is not a single model but a GPU-accelerated SDK that provides ASR, customizable LLM or NLP integration, and TTS as a cohesive platform (Introducing NVIDIA Riva: A GPU-Accelerated SDK for Developing Speech AI Applications | NVIDIA Technical Blog) . It’s a pipeline approach, but highly optimized for parallel processing on GPUs. Riva streams speech in, processes it through neural models, and streams out the response speech. Because it’s all running locally on GPU with optimized models, the pipeline achieves very low latency (even though it’s technically multiple models under the hood). Essentially, Riva offers building blocks to assemble a real-time voice assistant that feels integrated.

Latency: Far under 300 ms end-to-end on modern GPUs (Introducing NVIDIA Riva: A GPU-Accelerated SDK for Developing Speech AI Applications | NVIDIA Technical Blog) . In fact, Riva’s speech services can respond in ~150 ms on an NVIDIA A100 GPU (Riva Delivers Conversational AI on GPUs - NVIDIA Blog) . It supports true streaming ASR (producing interim transcripts with only tens of milliseconds delay) and fast-onset TTS. This enables near-instant voice agent responses. Many developers report achieving <500 ms round-trip latency with Riva for voice bots. (One reason is Riva’s efficient batching on GPUs and the ability to generate TTS audio concurrently with incoming text.) The low latency holds for reasonable utterance lengths, making it suitable for natural dialogue.

Languages & Accents: Riva supports multiple languages. Out-of-the-box, NVIDIA provides pre-trained models for English (US) and several other major languages (like Spanish, German, French, Russian, Mandarin, etc.), and they continually add more. The TTS component includes a selection of voices (you can also create custom voices with as little as 30 minutes of data (NVIDIA Riva Enables Enterprises to Create Unique Voices for Every ...) ). This means you can have different accents or genders for the synthesized voice as needed. ASR models are robust to various accents (e.g., the English model can handle global English accents reasonably, or you can choose locale-specific models). Additionally, because Riva allows custom model training, organizations can fine-tune models for particular accents or vocabulary if needed.

Deployment: Extremely flexible. Riva is designed for on-premises, edge, or private cloud deployment (Introducing NVIDIA Riva: A GPU-Accelerated SDK for Developing Speech AI Applications | NVIDIA Technical Blog) . It comes as Docker containers (through NVIDIA’s NGC catalog) that you can run on any machine with an NVIDIA GPU. This means you can deploy voice AI on an edge device (Jetson Orin for example) for on-device processing, or on your own servers for data privacy. It can also run in public cloud GPU instances if desired, but the key is you control the environment. This is ideal for scenarios where cloud services like GPT-4o or Azure can’t be used (due to latency of cloud hops or privacy regulations).

SDK/API: Yes – Riva provides gRPC and HTTP API endpoints for all its services. Developers use Riva SDK or REST calls to send audio and receive transcripts or send text and receive audio, etc. To create a full S2S loop, you’d typically connect Riva ASR → an LLM (could be on the same server, e.g. an open-source LLM or Nvidia’s NeMo models) → Riva TTS. NVIDIA has reference pipelines and even an AI workflow called ACE (Avatar Cloud Engine) that combines Riva with LLMs to build interactive avatars. So while not a single API call for the whole loop (you orchestrate a couple of calls), the SDK is built for real-time chaining. There are also client libraries to ease streaming audio to Riva and getting streaming responses.

Use Cases: Enterprise voice assistants, call center automation, automotive voice AI, healthcare dictation, and any application needing low-latency voice processing on secure infrastructure. Riva is used in systems where every millisecond counts – e.g., a robot that converses with you on the factory floor, or a VR game character reacting to player voice. Its on-prem nature also suits data-sensitive domains (like a hospital voice agent that must not send data to external clouds). Developers also leverage Riva to create custom voice experiences (since you can fine-tune voices and recognition for jargon). Essentially, it’s the toolkit for building your own “GPT-4o-like” voice agent when you can’t use a fully hosted API.

Comparison of Real-Time S2S Models

The table below summarizes key characteristics of each model/platform for quick reference:

Each of these solutions is suited for low-latency voice-driven applications, but they differ in maturity and availability. For cutting-edge integrated S2S with advanced reasoning, closed platforms like GPT-4o and Nova Sonic offer ready-to-use quality (at some cost, and cloud-only). Open-source projects like SeamlessM4T and LLaMA-Omni provide flexibility and community-driven innovation, letting developers run real-time voice models on their own hardware. And for those who still prefer a modular approach but need to minimize latency, toolkits like NVIDIA Riva demonstrate that optimized pipelines can achieve human-level response times (Introducing NVIDIA Riva: A GPU-Accelerated SDK for Developing Speech AI Applications | NVIDIA Technical Blog) .

In summary, the landscape is quickly evolving towards more natural, instant voice interactions, powered by these integrated models that blur the line between listening and speaking. Developers can choose a platform based on language needs, deployment constraints, and whether they prioritize an out-of-the-box service or an open model they can tweak. With options now ranging from big-tech APIs to open-source libraries, building a low-latency voice agent is more feasible than ever.

Sources: The information above is derived from official documentation and announcements for each model, including OpenAI’s GPT-4o release notes (Hello GPT-4o | OpenAI) (Hello GPT-4o | OpenAI) , AWS’s Nova Sonic launch blog (Introducing Amazon Nova Sonic: Human-like voice conversations for generative AI applications | AWS News Blog) (Introducing Amazon Nova Sonic: Human-like voice conversations for generative AI applications | AWS News Blog) , Microsoft’s Azure Voice Live preview docs (Voice Live API overview - Azure AI services | Microsoft Learn) , Meta’s SeamlessM4T research paper (SeamlessM4T-v2) (SeamlessM4T-v2) , the LLaMA-Omni GitHub repo (GitHub - ictnlp/LLaMA-Omni: LLaMA-Omni is a low-latency and high-quality end-to-end speech interaction model built upon Llama-3.1-8B-Instruct, aiming to achieve speech capabilities at the GPT-4o level.) , Google’s Translatotron research publication (Translatotron 2: Robust direct speech-to-speech translation) , and NVIDIA’s Riva technical blog (Introducing NVIDIA Riva: A GPU-Accelerated SDK for Developing Speech AI Applications | NVIDIA Technical Blog) (Introducing NVIDIA Riva: A GPU-Accelerated SDK for Developing Speech AI Applications | NVIDIA Technical Blog) . These sources provide detailed evidence of the capabilities, performance, and use cases discussed.