OpenAI's GPT-Live-1.1 Adds Multilingual Voices and Fixes Repetitive Tool-Call Audio
OpenAI's updated voice model adds new voices across six languages, more natural accents, and fewer awkward silences while searching.
- OpenAI released GPT-Live-1.1 and a mini variant called Quip for ChatGPT Voice
- New native voices added for French, German, Hindi, Japanese, Korean, and Mexican Spanish
- Wider range of filler phrases replaces the repetitive "checking..." during tool calls
- Available in ChatGPT, Codex, and dots today, API access coming soon
- Builds on full-duplex architecture that listens and speaks simultaneously
- Original GPT-Live-1 API pricing sits at $0.05 per minute with 12 voices
GPT-Live-1.1 adds new voices and smoother tool-call waits
OpenAI is rolling out GPT-Live-1.1, an incremental update to its real-time voice model. ChatGPT Voice gains new language and regional voice options, more natural accents, and a broader set of spoken cues for periods when the assistant is using tools. The update reaches all ChatGPT users starting today, according to the release notes, with API availability planned soon afterward.
New voices, fewer repeated cues
ChatGPT Voice receives GPT-Live-1.1 and a smaller variant called GPT-Live-1.1 Quip. OpenAI has not yet detailed the performance, latency, or pricing differences between the two models.
The initial voice additions cover French, German, Hindi, Japanese, Korean, and Mexican Spanish. OpenAI is also adding an English voice called Bubbie. Availability may vary by region and account, so developers should not assume every voice will appear in every deployment.
GPT-Live-1.1 also expands the phrases used while the model invokes tools such as web search or code execution. GPT-Live-1 often repeated a small set of cues, including “checking,” during longer calls. The update varies those responses, reducing repetitive audio while the tool completes its work.
| Surface | Availability |
|---|---|
| ChatGPT Voice | Rolling out to all users today |
| Codex and Dots | Included in the current rollout |
| OpenAI API | Coming later; no exact date announced |
Tool calls expose voice latency
Real-time voice applications need to keep users informed when a tool call interrupts the model’s response. Silence can resemble a dropped connection, while repeated status phrases make the assistant sound mechanical. GPT-Live-1.1 addresses the second problem by varying its spoken cues.
OpenAI’s notes do not claim lower tool latency. Search, code execution, and external API calls may still take the same amount of time. The change improves how the model handles that wait in conversation.
The Live architecture continuously processes incoming audio while generating speech. It can decide whether to continue speaking, pause, listen, respond to an interruption, or invoke a tool. The new status phrases refine the audible layer of that decision loop.
One model handles audio both ways
OpenAI’s earlier Advanced Voice Mode used a three-stage pipeline: speech recognition converted audio to text, a language model produced an answer, and text-to-speech generated the final audio. GPT-Live combines audio input and output in one model, allowing it to react during speech instead of waiting for a complete transcript and response.
That full-duplex design supports several behaviors relevant to application developers:
- Mid-sentence interruption and turn-taking
- Backchannel responses such as “mhmm” during user speech
- Tool and reasoning-model delegation within a conversation
- Phone-agent and customer-support workflows
Complex tasks can still depend on a separate reasoning model. The selected backend, tool chain, and network path therefore continue to affect response quality and end-to-end latency.
API details remain incomplete
OpenAI’s original API launch post listed GPT-Live-1 at $0.05 per minute with full-duplex speech, interruption handling, and 12 voice options. The company has not published GPT-Live-1.1 pricing, model identifiers, rate limits, or migration instructions.
Custom voices also remain sales-gated instead of self-service. OpenAI says the system is designed for conversational applications and includes safeguards against imitating a real person’s voice.
What developers should test
Teams already using GPT-Live-1 will need the API documentation before finalizing a migration. Once access opens, the main checks are operational:
- Confirm the exact model identifiers and endpoint compatibility.
- Compare interruption handling and turn-detection behavior with existing prompts.
- Measure tool-call latency separately from the new spoken status cues.
- Verify voice availability in every supported region and account tier.
- Review pricing and rate limits before moving production traffic.
The clearest gains apply to multilingual consumer apps, phone agents that make frequent tool calls, and voice-based Codex workflows. GPT-Live-1.1 broadens accent coverage and reduces repetitive waiting audio, while API cost, migration requirements, and performance differences remain pending.
Comments
