Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Fix OpenRouter MAI-Voice speech playback (#16463)

Daniel Han49aef9f762ea
diff --git a/studio/backend/core/inference/external_provider.py b/studio/backend/core/inference/external_provider.py
index 600a35b5497..710749394bb 100644
--- a/studio/backend/core/inference/external_provider.py
+++ b/studio/backend/core/inference/external_provider.py
@@ -7161,6 +7161,13 @@ class ExternalProviderClient:
"input": text,
"response_format": response_format,
}
+ mai_pcm = (
+ self.provider_type == "openrouter"
+ and model.strip().lower().startswith("microsoft/mai-voice-")
+ and response_format.strip().lower() == "wav"
+ )
+ if mai_pcm:
+ body["response_format"] = "pcm"
if voice:
body["voice"] = voice
if speed is not None:
@@ -7176,6 +7183,19 @@ class ExternalProviderClient:
response.raise_for_status()
media_type = (response.headers.get("content-type") or "").split(";")[0].strip()
audio = response.content
+ if mai_pcm:
+ if media_type.lower() not in ("audio/pcm", "application/octet-stream", ""):
+ raise httpx.DecodingError("OpenRouter MAI-Voice did not return PCM audio.")
+ if not audio or len(audio) % 2:
+ raise httpx.DecodingError("OpenRouter MAI-Voice returned malformed PCM audio.")
+ # https://openrouter.ai/docs/guides/overview/multimodal/tts#azure-mai-voice
+ output = io.BytesIO()
+ with wave.open(output, "wb") as writer:
+ writer.setnchannels(1)
+ writer.setsampwidth(2)
+ writer.setframerate(24_000)
+ writer.writeframes(audio)
+ return output.getvalue(), "audio/wav"
if response_format.strip().lower() == "wav":
merge_cancelled = threading.Event()
merge_task = asyncio.create_task(
diff --git a/studio/backend/tests/test_external_tts_formats.py b/studio/backend/tests/test_external_tts_formats.py
new file mode 100644
index 00000000000..1ee861bf7a1
--- /dev/null
+++ b/studio/backend/tests/test_external_tts_formats.py
@@ -0,0 +1,117 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+
+import asyncio
+import io
+import json
+import threading
+import wave
+from http.server import BaseHTTPRequestHandler, HTTPServer
+
+import httpx
+import pytest
+
+from core.inference import external_provider as provider_module
+from core.inference.external_provider import ExternalProviderClient
+
+
+_PCM = b"\x01\x00\xff\x7f\x00\x80"
+_MODEL = "microsoft/mai-voice-2.1-flash"
+_VOICE = "de-DE-Mia:MAI-Voice-2.1-Flash"
+
+
+def _speak(
+ monkeypatch,
+ provider,
+ model,
+ response_format,
+ accepted,
+ payload,
+ content_type,
+ repeat = 1,
+):
+ sent = []
+
+ class Handler(BaseHTTPRequestHandler):
+ def do_POST(self):
+ body = json.loads(self.rfile.read(int(self.headers["content-length"])))
+ sent.append(body)
+ allowed = body["response_format"] == accepted
+ self.send_response(200 if allowed else 400)
+ if content_type is not None:
+ self.send_header("content-type", content_type)
+ self.end_headers()
+ self.wfile.write(payload if allowed else b"response_format: expected mp3 or pcm")
+
+ def log_message(self, *_args):
+ pass
+
+ server = HTTPServer(("127.0.0.1", 0), Handler)
+ worker = threading.Thread(target = server.serve_forever, kwargs = {"poll_interval": 0.01})
+ worker.start()
+
+ async def run():
+ async with httpx.AsyncClient(trust_env = False) as transport:
+ monkeypatch.setattr(provider_module, "_http_client", transport)
+ client = ExternalProviderClient(
+ provider, f"http://127.0.0.1:{server.server_port}/v1", "test"
+ )
+ results = [
+ await client.create_speech(
+ text = "Hallo",
+ model = model,
+ voice = _VOICE,
+ response_format = response_format,
+ )
+ for _ in range(repeat)
+ ]
+ assert all(result == results[0] for result in results)
+ return results[0]
+
+ try:
+ result = asyncio.run(run())
+ finally:
+ server.shutdown()
+ worker.join()
+ server.server_close()
+ assert len(sent) == repeat
+ assert sent[0]["voice"] == _VOICE
+ assert sent[0]["input"] == "Hallo"
+ return result
+
+
[email protected]("content_type", ["audio/pcm", "Audio/PCM", "audio/pcm; rate=24000", None])
+def test_openrouter_mai_returns_playable_wav(monkeypatch, content_type):
+ audio, media_type = _speak(
+ monkeypatch, "openrouter", _MODEL, "wav", "pcm", _PCM, content_type, repeat = 2
+ )
+ assert media_type == "audio/wav"
+ with wave.open(io.BytesIO(audio), "rb") as reader:
+ assert reader.getframerate() == 24_000
+ assert reader.getnchannels() == 1
+ assert reader.getsampwidth() == 2
+ assert reader.readframes(reader.getnframes()) == _PCM
+
+
+ "provider, model, fmt, content_type",
+ [
+ ("custom", _MODEL, "wav", "audio/wav"),
+ ("openai", "gpt-4o-mini-tts", "wav", "audio/wav"),
+ ("openrouter", "openai/gpt-4o-mini-tts", "wav", "audio/wav"),
+ ("openrouter", _MODEL, "mp3", "audio/mpeg"),
+ ("openrouter", _MODEL, "pcm", "audio/pcm"),
+ ],
+)
+def test_other_requests_keep_format_and_bytes(monkeypatch, provider, model, fmt, content_type):
+ audio, media_type = _speak(monkeypatch, provider, model, fmt, fmt, _PCM, content_type)
+ assert audio == _PCM
+ assert media_type == content_type
+
+
+ "payload, content_type",
+ [(b"", "audio/pcm"), (b"\x01", "audio/pcm"), (b"ID3mp3", "audio/mpeg")],
+)
+def test_malformed_pcm_is_a_provider_http_error(monkeypatch, payload, content_type):
+ with pytest.raises(httpx.HTTPError):
+ _speak(monkeypatch, "openrouter", _MODEL, "wav", "pcm", payload, content_type)