From 61faad97e3a101b5aeb03756a7222d226bb9e826 Mon Sep 17 00:00:00 2001
From: octo-patch <266937838+octo-patch@users.noreply.github.com>
Date: Fri, 31 Jul 2026 04:19:05 +0000
Subject: [PATCH] Add MiniMax text-to-speech API tool for global and CN
 endpoints

---
 api.py              | 66 +++++++++++++++++++++++++++++++++++++++++++++
 examples.py         | 12 +++++++++
 gte.py              | 21 +++++++++++++++
 phi_3_vision_mlx.py |  2 +-
 4 files changed, 100 insertions(+), 1 deletion(-)

diff --git a/api.py b/api.py
index 10fc39d..675c1ed 100644
--- a/api.py
+++ b/api.py
@@ -1,8 +1,74 @@
 import os
 from pathlib import Path
 
+import requests
 from huggingface_hub import InferenceClient
 
+MINIMAX_TTS_DEFAULT_MODEL = "speech-2.8-hd"
+MINIMAX_TTS_MODELS = (
+    "speech-2.8-hd",
+    "speech-2.8-turbo",
+    "speech-2.6-hd",
+    "speech-2.6-turbo",
+    "speech-02-hd",
+    "speech-02-turbo",
+    "speech-01-hd",
+    "speech-01-turbo",
+)
+MINIMAX_TTS_AUDIO_FORMATS = ("mp3", "wav", "flac", "pcm")
+MINIMAX_TTS_BASE_URLS = {
+    "global_en": "https://api.minimax.io/v1/t2a_v2",
+    "cn_zh": "https://api.minimaxi.com/v1/t2a_v2",
+}
+
+def minimax_tts_api(prompt, model=MINIMAX_TTS_DEFAULT_MODEL, region="global_en",
+                    voice_setting=None, output_format="mp3", verbose=True,
+                    return_dict=True):
+    """
+    Synthesize speech from text through the MiniMax text-to-audio endpoint.
+
+    Both the global (api.minimax.io) and China (api.minimaxi.com) T2A endpoints
+    are supported via the ``region`` argument. The non-streaming response carries
+    the audio payload as a hex string in ``data.audio``; ``base_resp.status_code``
+    is ``0`` on success.
+
+    Example:
+    --------
+    agent = Agent(toolchain = "responses = minimax_tts_api(prompt)")
+    agent('People say nothing is impossible, but I do nothing every day.')
+    """
+    base_url = MINIMAX_TTS_BASE_URLS.get(region, MINIMAX_TTS_BASE_URLS["global_en"])
+    extension = output_format if output_format in MINIMAX_TTS_AUDIO_FORMATS else "mp3"
+    body = dict(
+        model=model,
+        text=prompt,
+        stream=False,
+        output_format=extension,
+    )
+    if voice_setting is not None:
+        body["voice_setting"] = voice_setting
+    response = requests.post(
+        f"{base_url}",
+        headers={
+            "Authorization": f"Bearer {os.environ.get('MINIMAX_API_KEY', '')}",
+            "Content-Type": "application/json",
+        },
+        json=body,
+    )
+    response.raise_for_status()
+    payload = response.json()
+    status_code = payload.get("base_resp", {}).get("status_code", -1)
+    if status_code != 0:
+        raise RuntimeError(f"MiniMax T2A request failed: {payload}")
+    audio_hex = payload.get("data", {}).get("audio")
+    audio_bytes = bytes.fromhex(audio_hex)
+    Path(f"minimax_tts.{extension}").write_bytes(audio_bytes)
+    if verbose:
+        print(f'### Prompt ###\n{prompt}\n### Saved ###\nminimax_tts.{extension}')
+    if return_dict:
+        return {'responses': prompt}
+    return prompt
+
 def mistral_api(prompt, history, verbose=True, return_dict=True, api_model="mistralai/Mistral-Nemo-Instruct-2407"):
     """
     Example:
diff --git a/examples.py b/examples.py
index c362375..91bf67c 100644
--- a/examples.py
+++ b/examples.py
@@ -63,6 +63,18 @@
 agent('Speak "People say nothing is impossible, but I do nothing every day."')
 agent.end()
 
+# Text-to-Speech
+
+## MiniMax TTS (speech-2.8-hd is the default; other MiniMax TTS models are also available)
+agent = pv.Agent(toolchain = "responses = minimax_tts_api(prompt)")
+agent('People say nothing is impossible, but I do nothing every day.')
+agent.end()
+
+### Selecting the China endpoint and a different MiniMax TTS model
+agent = pv.Agent(toolchain = 'responses = minimax_tts_api(prompt, region="cn_zh", model="speech-2.8-turbo")')
+agent('People say nothing is impossible, but I do nothing every day.')
+agent.end()
+
 # Toolchain
 
 ## LLM Backend Hotswap
diff --git a/gte.py b/gte.py
index 19560e7..163e4fc 100644
--- a/gte.py
+++ b/gte.py
@@ -173,6 +173,27 @@ def __call__(self, input_text: List[str]) -> mx.array:
 print('<|api_output|>'+result)
 ```
 """,
+"""Text to speech
+```python
+import os
+import requests
+endpoint = "https://api.minimax.io/v1/t2a_v2"
+response = requests.post(
+        endpoint,
+        headers={
+            "Authorization": "Bearer " + os.environ["MINIMAX_API_KEY"],
+            "Content-Type": "application/json",
+        },
+        json={
+            "model": "speech-2.8-hd",
+            "text": "{prompt}",
+            "stream": False,
+            "output_format": "mp3",
+        },
+)
+print('<|api_output|>'+response.json()["data"]["audio"])
+```
+""",
 """Transcribe youtube video
 ```python
 from gradio_client import Client
diff --git a/phi_3_vision_mlx.py b/phi_3_vision_mlx.py
index a8c9af7..b38beaa 100644
--- a/phi_3_vision_mlx.py
+++ b/phi_3_vision_mlx.py
@@ -26,7 +26,7 @@
 from mlx.utils import tree_flatten, tree_unflatten
 from PIL import Image
 
-from api import bark_api, mistral_api
+from api import bark_api, minimax_tts_api, mistral_api
 from gte import VDB, GteModel
 from phi import (LoRALinear, Phi3ForCausalLM, Phi3FProcessor, Phi3VForCausalLM,
                  Phi3VProcessor, Tic, TrainingCallback)