Speech 2.5 Turbo Preview

The Speech 2.5 API by MiniMax provides a high-fidelity, low-latency audio-native experience. It supports native speech-to-speech processing and 3-second zero-shot voice cloning, making it ideal for responsive, emotionally intelligent AI agents.

Modalities

—

Input

Text*

Voice_id*

Speed

Volume

Pitch

Emotion

English_normalization

This parameter supports English text normalization, which improves performance in number-reading scenarios.

Sample_rate

Bitrate

Channel

Format

Language_boost

Enable_sync_mode

If set to true, the function will wait for the result to be generated and uploaded before returning the response. It allows you to get the result directly in the response. This property is only available through the API.

Output

API Usage Examples

curl --request POST "https://gptproto.com/api/v3/minimax/speech-2.5-turbo-preview/text-to-audio" \
  --header "Authorization: Bearer $GPTPROTO_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "text": "A tiny origami fox sailing a teacup across a moonlit puddle",
    "voice_id": "Wise_Woman",
    "speed": 1,
    "volume": 1,
    "pitch": "0",
    "emotion": "happy",
    "english_normalization": false,
    "sample_rate": 8000,
    "bitrate": 32000,
    "channel": 1,
    "format": "mp3",
    "language_boost": "",
    "enable_sync_mode": false
  }'

Speech 2.5 Turbo Preview pricing

Chat, coding agents & document work. Priced per 1M tokens — input, cached input and output are billed separately. GPTProto is 40% below official rates.

Your usage

Model

MiniMax · ≈ 100M tokens/mo (0M cached)

Monthly volume

OpenRouter

List + 5.5% credit fee

$1,266

per month

Output$1,266

MiniMax

Direct from MiniMax (list price)

$1,200

per month

Output$1,200

BEST VALUE

GPTProto

40% off official

$720

per month

After discount$720

Effective$720

Monthly cost by source

OpenRouter

$1,266

MiniMax

$1,200

GPTProto

$720

Save $480 / month

on GPTProto vs MiniMax (−40%)

Annual savings ≈ $5,760 · pay as you go

GPTProto applies a per-model discount (10–30% off official) on top of bonus credits — every unit costs less than going direct.

Key Speech 2.5 API Features

Q: How does the Speech 2.5 API handle latency?

The Speech 2.5 API uses a native speech-to-speech (S2S) architecture that removes the intermediate text-transcription layer. This reduces the Time-To-First-Audio to approximately 280ms in optimal conditions. For the lowest possible latency in interactive applications, we recommend using the WebSocket streaming interface rather than standard chunked HTTP.

Q: Does this 2.5 model support voice cloning?

Yes, the 2.5 turbo model features zero-shot voice cloning. You can replicate a speaker's voice using just a 3-second audio sample. This process requires no secondary training or fine-tuning, allowing you to generate speech with a specific timbre or accent instantly via the /v1/audio/voices endpoint.

Q: What are the Speech 2.5 API pricing rates?

Pricing is based on both token count for input and audio duration for output. Input text is priced at $2.00 per 1M tokens, while input audio costs $0.015 per minute. Generated output audio is billed at $0.030 per minute. GPTProto.com users benefit from unified billing and 50% discounts on cached audio prompts for repeat sessions.

Q: Is my speech data used for model training?

No. Under our Enterprise Privacy Policy, all audio and text data processed through the Speech 2.5 API is not used for training by us or MiniMax. Your intellectual property and voice clones remain private and secure within your project environment.

Q: How does it compare to GPT-4o audio mode?

While GPT-4o has a slightly faster TTFA (~240ms), the Speech 2.5 API offers a higher sampling rate of 48kHz compared to GPT-4o's 24kHz. Additionally, Speech 2.5 provides superior non-verbal emotional cues, such as audible sighs and laughter, which are often limited in other low-latency models.

Q: Can I use the API for bilingual speech?

Absolutely. The model is optimized for bilingual code-switching across more than 25 languages. It can transition between languages like English and Chinese within a single sentence while keeping the speaker's identity and emotional delivery perfectly consistent.

Explore the technical capabilities that make the Speech 2.5 API the industry leader for emotional, real-time audio.

Native Speech-to-Speech

Direct audio-to-audio processing ensures a latency of <300ms for natural, fluid conversations.

Zero-Shot Voice Cloning

Replicate any voice with high fidelity using only a 3-second sample for instant personalization.

Dynamic Emotional Prosody

Generate laughter, sighs, and breathing sounds to create an incredibly realistic human presence.

48kHz HD Audio Output

Support for high-definition, studio-quality audio delivery suitable for professional production.

Speech 2.5 API Frequently Asked Questions

Get technical insights and pricing details for implementing the Speech 2.5 API in your real-time voice applications and services.

How does the Speech 2.5 API handle latency?

The Speech 2.5 API uses a native speech-to-speech (S2S) architecture that removes the intermediate text-transcription layer. This reduces the Time-To-First-Audio to approximately 280ms in optimal conditions. For the lowest possible latency in interactive applications, we recommend using the WebSocket streaming interface rather than standard chunked HTTP.

Does this 2.5 model support voice cloning?

Yes, the 2.5 turbo model features zero-shot voice cloning. You can replicate a speaker's voice using just a 3-second audio sample. This process requires no secondary training or fine-tuning, allowing you to generate speech with a specific timbre or accent instantly via the /v1/audio/voices endpoint.

What are the Speech 2.5 API pricing rates?

Pricing is based on both token count for input and audio duration for output. Input text is priced at $2.00 per 1M tokens, while input audio costs $0.015 per minute. Generated output audio is billed at $0.030 per minute. GPTProto.com users benefit from unified billing and 50% discounts on cached audio prompts for repeat sessions.

Is my speech data used for model training?

No. Under our Enterprise Privacy Policy, all audio and text data processed through the Speech 2.5 API is not used for training by us or MiniMax. Your intellectual property and voice clones remain private and secure within your project environment.

How does it compare to GPT-4o audio mode?

While GPT-4o has a slightly faster TTFA (~240ms), the Speech 2.5 API offers a higher sampling rate of 48kHz compared to GPT-4o's 24kHz. Additionally, Speech 2.5 provides superior non-verbal emotional cues, such as audible sighs and laughter, which are often limited in other low-latency models.

Can I use the API for bilingual speech?

Absolutely. The model is optimized for bilingual code-switching across more than 25 languages. It can transition between languages like English and Chinese within a single sentence while keeping the speaker's identity and emotional delivery perfectly consistent.