Last update: June 30, 2026
MiniMax Speech 2.6 is an earlier MiniMax text-to-speech and text-to-audio model family, announced on October 30, 2025, for low-latency voice-agent workflows, multilingual speech synthesis, and improved handling of non-standard text formats. It is relevant for teams that need to understand an existing Speech 2.6 workflow, compare the HD and Turbo variants, or build around exact model IDs such as speech-2.6-hd and speech-2.6-turbo. As of June 30, 2026, MiniMax’s current model documentation lists Speech 2.6 under Legacy Audio Models, while newer Speech 2.8 variants are available for current expressive TTS workflows.
This guide focuses on practical implementation. Instead of treating Speech 2.6 as a news item, it explains what the model does, how the MiniMax Speech 2.6 API is structured, where it fits in voice-agent systems, and what teams should test before production.
Quick Facts About MiniMax Speech 2.6
| Item | Details |
|---|---|
| Model family | MiniMax text-to-speech / text-to-audio |
| Current status | Legacy Speech 2.x model family in current MiniMax docs; still selectable in T2A HTTP docs, but newer Speech 2.8 HD and Turbo variants are the current expressive TTS options |
| Main model IDs | speech-2.6-hd, speech-2.6-turbo |
| Variants | MiniMax Speech 2.6 HD and MiniMax Speech 2.6 Turbo |
| Primary use cases | Voice agents, narration, support bots, AI companions, IVR, accessibility, product onboarding |
| Language support | 40 languages and specified languages or dialects |
| Emotion support | 7 emotions |
| HTTP endpoint | POST /v1/t2a_v2 |
| WebSocket path | wss://api.minimax.io/ws/v1/t2a_v2 |
| Streaming support | Supported through HTTP streaming behavior and WebSocket workflows |
| Best-fit workflows | Real-time voice interfaces, multilingual speech synthesis, cloned-voice content, and app-generated audio |
MiniMax documentation lists speech-2.6-hd and speech-2.6-turbo under audio Legacy Models, with Speech 2.6 HD positioned around similarity and ultra-high quality, and Speech 2.6 Turbo positioned around value and low latency. The same documentation lists 40 languages, 7 emotions, and support for specified languages and dialects for both variants.
What Is MiniMax Speech 2.6?
MiniMax Speech 2.6 is a TTS/T2A model family that converts written text into spoken audio. In practical terms, it takes text from an application, applies voice and audio settings, and returns generated speech that can be played, stored, streamed, or embedded into another product.
The model family is especially relevant for developers and product teams building voice-first software. A voice agent, for example, may use automatic speech recognition to understand the user, an LLM to plan the response, and MiniMax TTS to speak the answer back. In that chain, the TTS layer must be fast, intelligible, natural enough for repeated use, and reliable with messy text from dynamic systems.
Speech 2.6 can also support non-agent workflows. Teams can use it for e-learning narration, AI companion voices, multilingual onboarding videos, accessibility features, support scripts, internal training tools, and generated voiceovers for SaaS products.
Core Features of MiniMax Speech 2.6
Low-Latency Speech Generation
MiniMax positioned Speech 2.6 around an optimized audio generation pipeline with under 250 milliseconds of end-to-end latency in the provider’s described generation path. That matters for conversational interfaces because even a good voice can feel awkward if responses arrive too slowly.
For production planning, treat the model’s generation latency as one part of the full voice stack. User-perceived delay also depends on speech recognition, the LLM, external tools, network conditions, buffering, and playback logic.
Non-Standard Text Handling
A common TTS challenge is reading text that was not written for narration. Customer support systems often produce account balances, phone numbers, dates, email addresses, URLs, IP addresses, order IDs, and product codes. MiniMax’s Speech 2.6 announcement emphasized improved handling of non-standard formats such as URLs, email addresses, phone numbers, dates, monetary amounts, and IP addresses.
This is useful when the source text is generated by an LLM, CRM, support database, or checkout system. Even so, teams should test their own examples, especially names, abbreviations, domain terms, and mixed-language phrases.
Multilingual Speech Synthesis
Speech 2.6 HD and Turbo are documented with support for 40 languages and specified dialects. The language_boost parameter can help enhance recognition for selected languages and dialects, and it can be set to auto when the language type is unclear.
For multilingual products, the key implementation step is not only choosing the model. It is also testing scripts in the exact language mix used by customers, including names, local addresses, numerals, currency, and support phrases.
Emotion and Prosody Control
Speech 2.6 documentation lists support for 7 emotions. This can help product teams move beyond neutral narration into calmer support responses, warmer onboarding flows, or more expressive character speech. Emotion control should be used carefully: a support bot, for example, should sound calm and helpful rather than theatrical.
Prosody also matters. Pauses, speed, volume, pitch, and voice selection often shape user experience more than a single model choice. MiniMax supports pause markers such as <#0.4#> for timing control, along with voice and audio settings in API requests.
Voice Cloning and Fluent LoRA
MiniMax described Fluent LoRA as part of Speech 2.6’s cloned-voice improvements, focused on preserving timbre while producing more fluent and natural speech from imperfect or non-native source recordings. For actual cloning workflows, also review MiniMax’s clone-audio upload requirements before production use.
Voice cloning can be powerful, but it also adds legal, ethical, and operational risk. Use it only with clear consent, rights clearance, and a defined usage scope.
Streaming and Subtitle Support
Speech 2.6 can be used in workflows where streaming matters, especially conversational systems. MiniMax’s WebSocket guide demonstrates real-time speech synthesis using wss://api.minimax.io/ws/v1/t2a_v2. The HTTP T2A documentation also includes subtitle_enable for Speech 2.6 HD and Turbo, with sentence-level, word-level, and streaming-oriented word timestamps available through subtitle settings.
MiniMax Speech 2.6 HD vs Turbo
| Category | speech-2.6-hd | speech-2.6-turbo |
|---|---|---|
| Main goal | Higher similarity and ultra-high quality | Value and low latency |
| Quality focus | Better fit for polished narration, branded audio, and cloned-voice quality checks | Better fit for responsive voice interactions |
| Latency focus | Suitable when quality is the stronger priority | Suitable when speed and throughput matter more |
| Use cases | E-learning, product videos, voiceovers, premium assistant voices, careful voice cloning | Voice agents, support bots, IVR, interactive assistants, high-volume dynamic speech |
| Cost planning | May be chosen for fewer, higher-value outputs | May be chosen for larger volumes and real-time usage |
| Best fit | Experiences where final audio quality carries more weight | Experiences where response time and operating cost carry more weight |
Choose MiniMax Speech 2.6 HD when the listening experience is the main product surface: narrated lessons, brand voiceovers, sales demos, training audio, or cloned voices that need extra review. Choose MiniMax Speech 2.6 Turbo when the system must respond quickly and repeatedly, such as a voice agent, customer support assistant, interactive AI companion, or real-time product guide.
The right choice should come from testing. Use the same scripts, voice IDs, language mix, and output settings for both variants. Then compare naturalness, pronunciation, time to first audio, total response time, and cost under your expected usage pattern.
MiniMax Speech 2.6 API: How It Works
A typical MiniMax Speech 2.6 API workflow looks like this:
Text input → model selection → voice_id → voice settings → audio settings → streaming or file output.
For HTTP usage, MiniMax documents POST /v1/t2a_v2 for synchronous T2A. The required body fields include model and text, and the documented model options include speech-2.6-hd and speech-2.6-turbo. The text field must be under 10,000 characters, and streaming output is recommended for text over 3,000 characters.
Here is a clean non-streaming cURL example using speech-2.6-turbo:
curl --request POST \
--url https://api.minimax.io/v1/t2a_v2 \
--header "Authorization: Bearer $MINIMAX_API_KEY" \
--header "Content-Type: application/json" \
--data '{
"model": "speech-2.6-turbo",
"text": "Welcome to the product demo.<#0.4#> Your order total is $1,234.56, and your support contact is [email protected].",
"stream": false,
"language_boost": "auto",
"output_format": "url",
"voice_setting": {
"voice_id": "English_expressive_narrator",
"speed": 1,
"vol": 1,
"pitch": 0
},
"audio_setting": {
"sample_rate": 32000,
"bitrate": 128000,
"format": "mp3",
"channel": 1
}
}'Several request fields deserve extra attention. language_boost helps with language or dialect recognition and can be set to auto. audio_setting controls sample rate, bitrate, format, and channel. subtitle_enable can be used when timestamped subtitle output is needed. For non-streaming output, output_format can be url or hex; in MiniMax’s HTTP documentation, returned URLs are valid for 24 hours.
Pronunciation controls should also be part of your QA process. The API documentation supports pause markers in the form <#x#> and inline pronunciation overrides using Mandarin Pinyin, IPA, or Cantonese Jyutping in parentheses. Do not rely on interjection tags such as laughs, coughs, breaths, or sighs in Speech 2.6 workflows; MiniMax documents those tags for Speech 2.8 HD and Turbo.
Practical Use Cases
MiniMax Speech 2.6 is a strong fit for voice agents and customer support tools because it combines low-latency design with text handling that is useful for account data, dates, phone numbers, and emails. A support bot can read a balance, confirm a case number, or explain the next step without forcing the engineering team to normalize every string manually.
AI companions and character-based products can use MiniMax voice cloning and emotion settings to create more consistent personalities. This is especially useful when the product experience depends on a recognizable voice style, but every cloned voice should be approved and governed.
E-learning teams can use Speech 2.6 HD for training narration, short lessons, pronunciation aids, and multilingual content. Product teams can use Turbo for onboarding flows that respond to user behavior inside an application. Accessibility tools can turn written interfaces into speech for users who prefer or need audio output.
Other use cases include IVR systems, internal SaaS assistants, podcast-style summaries, video explainers, product demos, and automated status updates.
Pricing and Cost Planning
MiniMax Speech 2.6 pricing should be treated as cost planning, not as a permanent pricing table. MiniMax’s official pay-as-you-go pricing currently lists legacy audio pricing at $60/M characters for speech-2.6-turbo and $100/M characters for speech-2.6-hd. Pricing can change, and third-party providers may package Speech 2.6 differently, so verify the official MiniMax pay-as-you-go pricing before production planning.
The main cost drivers are monthly generated characters, HD vs Turbo selection, direct API access vs marketplace access, streaming or async workflow design, voice cloning or voice design fees, retries from failed generations, and storage or processing for generated audio files.
A simple budgeting formula is:
Monthly generated characters × selected character rate + voice-related fees + storage/processing costs
For production planning, estimate normal usage, peak usage, failed-request retries, and QA re-generation. If the product uses long-form narration, also plan for segmentation, review passes, and file storage. If the product uses a voice agent, model cost should be evaluated alongside ASR, LLM, tool calls, infrastructure, and telephony.
Limitations and Implementation Checks
Speech quality should be evaluated with real scripts, not generic samples. Test customer names, product names, addresses, order numbers, phone numbers, URLs, email addresses, dates, amounts, acronyms, code-like strings, and mixed-language text.
Total app latency is not only a TTS metric. Voice-agent delay can come from ASR capture, silence detection, LLM reasoning, tool calls, network hops, audio generation, buffering, and playback. Measure the full path from user speech to audible response.
Long text should be segmented, streamed, or handled through an async workflow. The synchronous T2A HTTP text field has a documented limit under 10,000 characters, and MiniMax recommends streaming for text over 3,000 characters. For longer narration, review MiniMax’s asynchronous long-form TTS workflow and design a chunking strategy that respects paragraph boundaries and preserves natural pacing.
Provider availability can vary. A direct MiniMax integration may expose different controls than a hosted marketplace or reseller. Confirm support for your required voice IDs, emotion settings, audio format, subtitles, streaming mode, and output format before committing to a production architecture.
MiniMax’s current rate-limit documentation lists T2A for speech-2.6-turbo and speech-2.6-hd at 60 RPM, alongside Speech 2.8 and Speech-02 T2A routes. Account-level limits, provider wrappers, and enterprise arrangements can still differ.
Also confirm output handling. If you request URL output, do not treat returned links as permanent storage. Download, store, or cache audio according to your product’s compliance and retention requirements.
Responsible Voice Cloning With MiniMax Speech 2.6
Responsible voice cloning starts with permission. Use only approved voices, obtain clear consent, and document who can use the cloned voice, where it can be used, and for how long.
Protect cloned voice_id values as sensitive production assets. Restrict access, audit usage, and avoid exposing them in client-side code. For commercial content, review generated audio before publication, especially when the voice represents a person, brand, or character.
Avoid impersonation, deception, or unauthorized recreation of a person’s voice. A safe workflow should include rights clearance, usage logs, human review for high-impact content, and a process for revoking or replacing cloned voices.
MiniMax Speech 2.6 vs Speech 2.8 and Earlier Speech Models
MiniMax Speech 2.6 remains useful for teams with existing speech-2.6-hd or speech-2.6-turbo workflows, especially where legacy pricing, voice compatibility, or production stability matters. However, MiniMax’s current documentation lists Speech 2.6 under Legacy Audio Models, while Speech 2.8 is the newer family for expressive TTS workflows with sound and interjection tags.
If you are choosing a model for a new integration, compare Speech 2.6 against MiniMax Speech 2.8, MiniMax Speech-02, and MiniMax Speech 2.5. Use the same scripts, languages, voices, output formats, and latency targets before deciding.
MiniMax Speech 2.6 FAQ
What is MiniMax Speech 2.6?
MiniMax Speech 2.6 is a MiniMax text-to-speech and text-to-audio model family used to convert written text into spoken audio for applications such as voice agents, narration, support bots, and multilingual product experiences.
What is the difference between MiniMax Speech 2.6 HD and Turbo?
MiniMax Speech 2.6 HD is better suited to higher-quality and similarity-focused outputs, while MiniMax Speech 2.6 Turbo is better suited to low-latency and value-oriented workflows. In practice, HD fits polished narration and cloned-voice review, while Turbo fits interactive voice agents and high-volume dynamic speech.
What are the model IDs for MiniMax Speech 2.6?
The main model IDs are speech-2.6-hd and speech-2.6-turbo. These are the values developers use in the model field when calling the MiniMax T2A API.
Is MiniMax Speech 2.6 good for voice agents?
Yes, it can be a good fit for voice agents because Speech 2.6 emphasizes low-latency speech generation, multilingual support, and handling of dynamic text formats such as dates, phone numbers, emails, and monetary amounts. Teams should still test the full voice pipeline, not only the TTS request.
Does MiniMax Speech 2.6 support voice cloning?
Yes, Speech 2.6 can be used in voice cloning workflows, and MiniMax described Fluent LoRA as part of its cloned-voice fluency improvements. Voice cloning should be used only with consent, rights clearance, and safeguards against misuse.
How many languages does MiniMax Speech 2.6 support?
MiniMax documentation lists 40 languages and support for specified languages and dialects for speech-2.6-hd and speech-2.6-turbo.
What API endpoint does MiniMax Speech 2.6 use?
For synchronous HTTP T2A, MiniMax documents POST /v1/t2a_v2. For WebSocket workflows, MiniMax’s guide uses wss://api.minimax.io/ws/v1/t2a_v2.
How should I estimate MiniMax Speech 2.6 cost?
Estimate cost by calculating monthly generated characters, provider character rate, voice-related fees, storage, retries, and processing costs. MiniMax’s current pay-as-you-go pricing lists speech-2.6-turbo at $60/M characters and speech-2.6-hd at $100/M characters under Legacy Audio Models, but prices can change and third-party providers may charge differently.
Conclusion
MiniMax Speech 2.6 remains useful as a legacy version-specific reference point for teams that need exact model IDs, multilingual speech synthesis, low-latency voice-agent workflows, voice cloning considerations, and practical T2A API integration around speech-2.6-hd or speech-2.6-turbo.
The safest production approach is to test both speech-2.6-hd and speech-2.6-turbo with your real scripts, voices, languages, formats, and latency targets, then compare them with current Speech 2.8 options before new production adoption. Choose the variant that matches the product experience, pricing, latency, voice compatibility, and review workflow, not only the model description.
More on speech: see the MiniMax Audio hub for voices, voice cloning, text-to-speech pricing and which speech model to choose.