#speechtospeech — Public Fediverse posts

OpenAI Realtime против Яндекс Realtime: сравнил два голосовых движка для России и за вечер пересадил свой прод

6 июля OpenAI выпустил gpt-realtime — движок «речь в речь», который говорит как человек и держит перебивание. Я делаю голосового ИИ-приёмщика для бизнеса на российском стеке и занервничал: не устарел ли мой подход? За вечер собрал прототип на OpenAI, откопал, что у Яндекса есть свой Realtime, и замерил задержку у всех трёх подходов одним методом. Цифры честные, с оговоркой про VPN: Яндекс Realtime ~330 мс, OpenAI ~740 мс (под VPN — иначе в РФ никак), старая трёхзвенка ~4,4 секунды. В статье — грабли обоих движков (400-й эндпоинт у OpenAI, три роли доступа и закрытый обмен OAuth→IAM у Яндекса), почему у OpenAI не лечится акцент в русском, и как я за вечер переписал свой прод на Яндекс Realtime: WebSocket-мост, живой PCM-плеер вместо MP3, перебивание без наушников. С кодом и реальными замерами.

https://habr.com/ru/articles/1057176/

#OpenAI_Realtime #Яндекс_SpeechKit #голосовой_ассистент #webrtc #speechtospeech

OpenAI Realtime против Яндекс Realtime: сравнил два голосовых движка для России и за вечер пересадил свой прод

6 июля OpenAI выпустил gpt-realtime — движок «речь в речь», который говорит как человек и держит перебивание. Я делаю голосового ИИ-приёмщика для бизнеса на российском стеке и занервничал: не устарел ли мой подход? За вечер собрал прототип на OpenAI, откопал, что у Яндекса есть свой Realtime, и замерил задержку у всех трёх подходов одним методом. Цифры честные, с оговоркой про VPN: Яндекс Realtime ~330 мс, OpenAI ~740 мс (под VPN — иначе в РФ никак), старая трёхзвенка ~4,4 секунды. В статье — грабли обоих движков (400-й эндпоинт у OpenAI, три роли доступа и закрытый обмен OAuth→IAM у Яндекса), почему у OpenAI не лечится акцент в русском, и как я за вечер переписал свой прод на Яндекс Realtime: WebSocket-мост, живой PCM-плеер вместо MP3, перебивание без наушников. С кодом и реальными замерами.

https://habr.com/ru/articles/1057176/

#OpenAI_Realtime #Яндекс_SpeechKit #голосовой_ассистент #webrtc #speechtospeech

OpenAI Realtime против Яндекс Realtime: сравнил два голосовых движка для России и за вечер пересадил свой прод

6 июля OpenAI выпустил gpt-realtime — движок «речь в речь», который говорит как человек и держит перебивание. Я делаю голосового ИИ-приёмщика для бизнеса на российском стеке и занервничал: не устарел ли мой подход? За вечер собрал прототип на OpenAI, откопал, что у Яндекса есть свой Realtime, и замерил задержку у всех трёх подходов одним методом. Цифры честные, с оговоркой про VPN: Яндекс Realtime ~330 мс, OpenAI ~740 мс (под VPN — иначе в РФ никак), старая трёхзвенка ~4,4 секунды. В статье — грабли обоих движков (400-й эндпоинт у OpenAI, три роли доступа и закрытый обмен OAuth→IAM у Яндекса), почему у OpenAI не лечится акцент в русском, и как я за вечер переписал свой прод на Яндекс Realtime: WebSocket-мост, живой PCM-плеер вместо MP3, перебивание без наушников. С кодом и реальными замерами.

https://habr.com/ru/articles/1057176/

#OpenAI_Realtime #Яндекс_SpeechKit #голосовой_ассистент #webrtc #speechtospeech

Как устроены voice-2-voice модели

Речь входит — речь выходит. Что внутри voice-2-voice моделей: как нейронный кодек превращает звук в токены, почему каскад ASR → LLM → TTS проигрывает миллисекунды и как Moshi умудряется поддакивать, пока вы говорите. Без магии, с примерами на Swift.

https://habr.com/ru/articles/1056820/

#голосовые_ассистенты #voicetovoice #speechtospeech #распознавание_речи #синтез_речи #LLM #нейронные_аудиокодеки #OpenAI_Realtime_API #Gemini_Live #moshi

The #SpeechToSpeech project was deployed. And it worked 🎉

It stayed live in one of the #PyConIT 2026 rooms throughout the event 🏗️

📊 A few words on the stats: the most used languages were English, French, Italian and Spanish. And someone even tried Korean 🤭

Thank you to #PyConIT 2026 for hosting my #talk

https://github.com/bilardi/realtime-speech-to-speech

#TalksAndWorkshops #aws #polly #sts #python

The #SpeechToSpeech project was deployed. And it worked 🎉

It stayed live in one of the #PyConIT 2026 rooms throughout the event 🏗️

📊 A few words on the stats: the most used languages were English, French, Italian and Spanish. And someone even tried Korean 🤭

Thank you to #PyConIT 2026 for hosting my #talk

https://github.com/bilardi/realtime-speech-to-speech

#TalksAndWorkshops #aws #polly #sts #python

The #SpeechToSpeech project was deployed. And it worked 🎉

It stayed live in one of the #PyConIT 2026 rooms throughout the event 🏗️

📊 A few words on the stats: the most used languages were English, French, Italian and Spanish. And someone even tried Korean 🤭

Thank you to #PyConIT 2026 for hosting my #talk

https://github.com/bilardi/realtime-speech-to-speech

#TalksAndWorkshops #aws #polly #sts #python

The #SpeechToSpeech project was deployed. And it worked 🎉

It stayed live in one of the #PyConIT 2026 rooms throughout the event 🏗️

📊 A few words on the stats: the most used languages were English, French, Italian and Spanish. And someone even tried Korean 🤭

Thank you to #PyConIT 2026 for hosting my #talk

https://github.com/bilardi/realtime-speech-to-speech

#TalksAndWorkshops #aws #polly #sts #python

Ho preparato un progetto #SpeechToSpeech che va in produzione alla #PyConIT 2026, attivo tutto il giorno in una sala.

Cos'ha di speciale ? È il primo a usare #AmazonPolly in #BidirectionalStreaming interamente in #python 🐍.

🎤 A #PyConIT 2026 il mio #talk: dalle necessità alle scelte, fino ai grattacapi affrontati.

https://2026.pycon.it/it/event/when-boto3-doesnt-have-it-yet-you-write-it-a-real-time-speech-to-speech-story-in-python

#TalksAndWorkshops #aws #polly #sts

Ho preparato un progetto #SpeechToSpeech che va in produzione alla #PyConIT 2026, attivo tutto il giorno in una sala.

Cos'ha di speciale ? È il primo a usare #AmazonPolly in #BidirectionalStreaming interamente in #python 🐍.

🎤 A #PyConIT 2026 il mio #talk: dalle necessità alle scelte, fino ai grattacapi affrontati.

https://2026.pycon.it/it/event/when-boto3-doesnt-have-it-yet-you-write-it-a-real-time-speech-to-speech-story-in-python

#TalksAndWorkshops #aws #polly #sts

It all started from a curiosity .. and from a new feature.

In a realtime #SpeechToSpeech project on #AWS, can I use #polly bidirectional streaming with #python ?

🔮 Spoiler: well, when boto3 doesn't have it (yet), you write it 🏗️

In the article I describe each choice and the reasons, eager to ship something new 😄

https://alessandra.bilardi.net/diary/articles/2026-05/when-boto3-doesn-t-have-it-yet-you-write-it-a-realtime-speech-to-speech-story-in-python.en

#DiaryOfALazyDeveloper #docker #fastapi

It all started from a curiosity .. and from a new feature.

In a realtime #SpeechToSpeech project on #AWS, can I use #polly bidirectional streaming with #python ?

🔮 Spoiler: well, when boto3 doesn't have it (yet), you write it 🏗️

In the article I describe each choice and the reasons, eager to ship something new 😄

https://alessandra.bilardi.net/diary/articles/2026-05/when-boto3-doesn-t-have-it-yet-you-write-it-a-realtime-speech-to-speech-story-in-python.en

#DiaryOfALazyDeveloper #docker #fastapi

It all started from a curiosity .. and from a new feature.

In a realtime #SpeechToSpeech project on #AWS, can I use #polly bidirectional streaming with #python ?

🔮 Spoiler: well, when boto3 doesn't have it (yet), you write it 🏗️

In the article I describe each choice and the reasons, eager to ship something new 😄

https://alessandra.bilardi.net/diary/articles/2026-05/when-boto3-doesn-t-have-it-yet-you-write-it-a-realtime-speech-to-speech-story-in-python.en

#DiaryOfALazyDeveloper #docker #fastapi

It all started from a curiosity .. and from a new feature.

In a realtime #SpeechToSpeech project on #AWS, can I use #polly bidirectional streaming with #python ?

🔮 Spoiler: well, when boto3 doesn't have it (yet), you write it 🏗️

In the article I describe each choice and the reasons, eager to ship something new 😄

https://alessandra.bilardi.net/diary/articles/2026-05/when-boto3-doesn-t-have-it-yet-you-write-it-a-realtime-speech-to-speech-story-in-python.en

#DiaryOfALazyDeveloper #docker #fastapi

Nvidia PersonaPlex 7B on Apple Silicon: Full-Duplex Speech-to-Speech in Swift

https://blog.ivan.digital/nvidia-personaplex-7b-on-apple-silicon-full-duplex-speech-to-speech-in-native-swift-with-mlx-0aa5276f2e23

#HackerNews #Nvidia #Apple #Silicon #SpeechToSpeech #Swift #AI #Technology

Nvidia PersonaPlex 7B on Apple Silicon: Full-Duplex Speech-to-Speech in Swift

https://blog.ivan.digital/nvidia-personaplex-7b-on-apple-silicon-full-duplex-speech-to-speech-in-native-swift-with-mlx-0aa5276f2e23

#HackerNews #Nvidia #Apple #Silicon #SpeechToSpeech #Swift #AI #Technology

High-Fidelity Simultaneous Speech-to-Speech Translation

https://arxiv.org/abs/2502.03382

#HackerNews #HighFidelity #Translation #SimultaneousSpeech #SpeechToSpeech #Technology

High-Fidelity Simultaneous Speech-to-Speech Translation

https://arxiv.org/abs/2502.03382

#HackerNews #HighFidelity #Translation #SimultaneousSpeech #SpeechToSpeech #Technology

🔥 Announcing #HuggingFace's latest #SpeechToSpeech development! 🔥

🚀 For those looking for #LowLatency without the hassle of #ServerSetup or #CloudComputing issues - there's a solution!
💰 Check out their new #BlogPost, showing how to use #HuggingFace's #InferenceEndpoints to deliver ultra-low latency on an #NVIDIAL4 #GPU—for just $0.80/hour!
🛠️ The team created a custom #Docker image for low latency, and they're #OpenSource-ing the entire solution for everyone to use!
🎥 The video shows a #WordGame played against #Llama3 8B; the #latency is so low that the game flows seamlessly!

👉 Read all about how they did it in their blog post:

https://huggingface.co/blog/s2s_endpoint

💻 Want to try it yourself? Here's the code to get started: https://github.com/huggingface/speech-to-speech/pull/118 🚀

#AI #MachineLearning #NLP #ArtificialIntelligence #TechNews #DevOps #CloudInfrastructure

🔥 Announcing #HuggingFace's latest #SpeechToSpeech development! 🔥

🚀 For those looking for #LowLatency without the hassle of #ServerSetup or #CloudComputing issues - there's a solution!
💰 Check out their new #BlogPost, showing how to use #HuggingFace's #InferenceEndpoints to deliver ultra-low latency on an #NVIDIAL4 #GPU—for just $0.80/hour!
🛠️ The team created a custom #Docker image for low latency, and they're #OpenSource-ing the entire solution for everyone to use!
🎥 The video shows a #WordGame played against #Llama3 8B; the #latency is so low that the game flows seamlessly!

👉 Read all about how they did it in their blog post:

https://huggingface.co/blog/s2s_endpoint

💻 Want to try it yourself? Here's the code to get started: https://github.com/huggingface/speech-to-speech/pull/118 🚀

#AI #MachineLearning #NLP #ArtificialIntelligence #TechNews #DevOps #CloudInfrastructure

"This new [Copilot app for Android & iOS] feature offers similar functionality in terms of speech-to-speech, being able to interrupt it and have it reflect your vocal tone and emotions as Advanced Voice, but it was more casual and felt more natural, slightly less stilted than OpenAI's offering."

Microsoft Copilot Voice is more human-like than ChatGPT — and it's free to all users - Tom's Hardware
https://www.tomsguide.com/ai/copilot/microsoft-copilot-voice-is-more-human-like-than-chatgpt-and-its-free-to-all-users
#ai #copilot #microsoft #msftadvocate #openai #voice #speechtospeech

"This new [Copilot app for Android & iOS] feature offers similar functionality in terms of speech-to-speech, being able to interrupt it and have it reflect your vocal tone and emotions as Advanced Voice, but it was more casual and felt more natural, slightly less stilted than OpenAI's offering."

Microsoft Copilot Voice is more human-like than ChatGPT — and it's free to all users - Tom's Hardware
https://www.tomsguide.com/ai/copilot/microsoft-copilot-voice-is-more-human-like-than-chatgpt-and-its-free-to-all-users
#ai #copilot #microsoft #msftadvocate #openai #voice #speechtospeech