Streaming Speech

Use TTS 1.5 Flash for low-latency streaming speech generation. This capability uses the same endpoint as single-speaker speech and is enabled with version=flash-20260626 and stream=true.

Authentication

curl https://api.mosi.cn/v1/audio/speech \
  -H "Authorization: Bearer $MOSS_API_KEY"

Supported models

Activation requirements and constraints

• Pass the exact lowercase value model=moss-tts; letter case is not normalized.
• version is required and must be flash-20260626.
• stream is required and must be true.
• response_format is required; currently only pcm is supported.
• stream_format=sse returns SSE, with each frame formatted as data: JSON Event.
• Omitting stream_format or passing audio returns raw PCM bytes.
• sample_rate is not a public request field; passing a non-zero value returns a parameter error before task creation.
• stream=true cannot be combined with async=true, and streaming calls do not accept a non-empty webhook_url.
• voice_url, voice_data, and file_id are not supported for streaming calls; use voice_id to specify a voice ID.
• delivery_method and aigc_metadata currently have no effect on the streaming path; do not pass them.

Request fields

Request examples

curl --no-buffer -X POST "https://api.mosi.cn/v1/audio/speech" \
  -H "Authorization: Bearer $MOSS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "moss-tts",
    "version": "flash-20260626",
    "input": "Welcome to the Mossland API.",
    "voice_id": "<voice_id>",
    "language": "en",
    "speed": 1,
    "expected_duration_sec": 3,
    "stream": true,
    "response_format": "pcm",
    "stream_format": "sse"
  }'

curl --no-buffer -X POST "https://api.mosi.cn/v1/audio/speech" \
  -H "Authorization: Bearer $MOSS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "moss-tts",
    "version": "flash-20260626",
    "input": "Welcome to the Mossland API.",
    "voice_id": "<voice_id>",
    "stream": true,
    "response_format": "pcm"
  }' \
  --output speech.pcm

Streaming response

stream_format=sse returns SSE, with each data: line followed by a JSON Event. type is the event discriminator; clients should use it to select the corresponding field structure. Omitting stream_format or passing audio returns raw PCM bytes.

Some access paths return task.created first, while others start directly with speech.created. Clients must use each frame's type and must not depend on a fixed first event.

SSE Event Types

task.created

speech.created

speech.audio.delta

speech.audio.done

error

Raw PCM response headers

Response example

# SSE response example: pass stream_format=sse in the request
Content-Type: text/event-stream; charset=utf-8

# Some access paths omit task.created and start directly with speech.created
data: {"type":"task.created","task_id":"task-123","object":"audio.speech","status":"PROCESSING","model":"moss-tts"}

data: {"type":"speech.created","format":"pcm","sample_rate":48000,"channels":1,"bit_depth":16}

data: {"type":"speech.audio.delta","audio":"BASE64_PCM_CHUNK"}

data: {"type":"speech.audio.done"}

# Raw PCM response example: omit stream_format or pass stream_format=audio
Content-Type: audio/pcm
X-Sample-Rate: 48000
X-Channels: 1
X-Bit-Depth: 16

<raw pcm bytes>

Errors and disconnect semantics

• missing_required_field: input is missing or empty after trimming.
• unsupported_response_format / response_format: the streaming request does not explicitly use pcm, or passes another audio format.
• unsupported_stream / stream: stream_format is not audio/sse, or the resolved model does not support streaming.
• unsupported_for_stream / sample_rate: the streaming request passes a non-zero sample_rate.
• invalid_range / speed: speed is outside the 0.25-4 range.
• unsupported_with_async / stream: stream=true and async=true are passed together.
• unsupported_with_webhook / stream: the streaming request passes a non-empty webhook_url.
• Request validation, model resolution, streaming-capability checks, and task-creation errors occur before the response starts and return a standard JSON error.
• If SSE fails before speech.audio.done, the server sends an error data frame and closes the connection.
• If raw PCM fails after response headers or audio bytes have been written, the server can only close the connection; returned PCM may be incomplete.
• A client disconnect ends the current HTTP stream. Depending on the access path, it may cancel the upstream request or only the current subscription; clients must not assume that generation either always continues or is always canceled.
• In the raw PCM mode, EOF alone does not prove that the audio is complete.

Next steps