/v3/audio/speech ignores response_format and always returns 32-bit float WAV #4613

Describe the bug
The OpenAI-compatible text-to-speech endpoint /v3/audio/speech ignores response_format. For wav, pcm, mp3 and flac it returns the same body: a WAV file with 32-bit IEEE float samples (format tag 3). The response is also sent with Content-Type: application/json; charset=utf-8.

The OpenAI API specifies 16-bit PCM for wav and headerless 16-bit little-endian PCM for pcm, so clients written against it cannot play the output. For example, wyoming_openai (the Wyoming bridge used by Home Assistant) cannot parse the float WAV header and plays static: roryeckel/wyoming_openai#72

To Reproduce

1. Models repository: OpenVINO/Kokoro-82M-int8-ov pulled from Hugging Face into /models/OpenVINO/Kokoro-82M-int8-ov, graph generated with ovms --configure --model_path /models/OpenVINO/Kokoro-82M-int8-ov --target_device CPU:

   input_stream: "HTTP_REQUEST_PAYLOAD:input"
   output_stream: "HTTP_RESPONSE_PAYLOAD:output"
   node {
       name: "T2sExecutor"
       calculator: "T2sCalculator"
       input_side_packet: "TTS_NODE_RESOURCES:t2s_servable"
       input_stream: "HTTP_REQUEST_PAYLOAD:input"
       output_stream: "HTTP_RESPONSE_PAYLOAD:output"
       node_options: {
           [type.googleapis.com / mediapipe.T2sCalculatorOptions]: {
               models_path: "/models/OpenVINO/Kokoro-82M-int8-ov"
               target_device: "CPU"
               plugin_config: '{"NUM_STREAMS":"1"}'
               }
       }
   }
   

2. OVMS launch command (container openvino/model_server:latest-gpu):

   --rest_port 8080 --config_path /config/config.json --cache_dir /cache --log_level INFO
   

3. Client command:

   for rf in wav pcm mp3 flac; do
     curl -s http://localhost:8080/v3/audio/speech -H 'Content-Type: application/json' \
       -d "{\"model\":\"OpenVINO/Kokoro-82M-int8-ov\",\"input\":\"Hello there.\",\"voice\":\"am_adam\",\"response_format\":\"$rf\"}" \
       -o "out.$rf"
     file -b "out.$rf"
   done

4. Every file is identical (146444 bytes):

   RIFF (little-endian) data, WAVE audio, IEEE Float, mono 24000 Hz
   

   Response headers:

   HTTP/1.1 200 OK
   content-length: 146444
   content-type: application/json; charset=utf-8
   

Expected behavior

• wav (the default) returns a 16-bit PCM WAV, as OpenAI does.
• pcm returns headerless 16-bit little-endian PCM.
• Formats OVMS cannot produce (e.g. mp3, flac) are rejected with a 400 error instead of silently returning a WAV.
• The Content-Type matches the audio format (e.g. audio/wav).

Logs
Nothing is logged for these requests at --log_level INFO; I have not captured DEBUG logs.

Configuration

1. OVMS version: OpenVINO Model Server 2026.4.0.869b2186a, OpenVINO backend 2026.4.0-22959-99c81491cc3-releases/2026/4, OpenVINO GenAI backend 2026.4.0.0-3407-7ea2546852a
2. config.json:

   {
     "model_config_list": [
       { "config": { "name": "OpenVINO/Kokoro-82M-int8-ov", "base_path": "/models/Kokoro-82M-int8-ov-cpu" } }
     ]
   }

3. Intel Core Ultra 5 125H (Meteor Lake), target device CPU
4. Model repository:

   /models/Kokoro-82M-int8-ov-cpu/graph.pbtxt
   /models/OpenVINO/Kokoro-82M-int8-ov/{config.json, openvino_config.json, openvino_model.xml, openvino_model.bin, voices/, ...}
   

5. Model: https://huggingface.co/OpenVINO/Kokoro-82M-int8-ov