When a CosyVoice 3 Flash emotion instruction does not produce the expected delivery, check the selected voice's Instruct support before changing the prompt. Then check the spoken language, instruction format and service configuration. An accepted instruction field does not establish that every voice can follow it.
Check why emotion is missing: Editorial workflow diagram, not a tested model output
Start with the exact voice
The official voice list, updated October 8, 2026, lists language, Instruct, SSML and timestamp capabilities by voice.
For example, longanyang supports Instruct with a defined format, while longanlang_v3 is listed without Instruct support. Speech synthesis support alone does not imply the same emotion controls.
Check the entry under the exact model and voice identifier instead of relying on a similar voice name from another version.
Identify the failing layer with a short comparison
Create a two-sentence baseline without an instruction, then add one supported emotion while keeping the voice and text fixed. This is a diagnostic method, not a measured fix rate.
Symptom: Successful task, little emotional change · First check: Instruct support for that exact voice · Next action: Compare a clearly supported voice using the same text
Symptom: Incorrect pronunciation or accent · First check: Languages supported by the selected voice · Next action: Verify plain reading before adding emotion
Symptom: Rejected request · First check: Mode, voice/voice_id, instruction length and required fields · Next action: Keep the task ID/error and correct the documented input
Symptom: Unavailable cloned voice · First check: Creation state, model/user binding and service region · Next action: Check the feature-specific restriction instead of declaring a country ban
Symptom: Access failure on one network or in one country · First check: Network, authentication or permission error · Next action: Separate platform connectivity from provider region resources
For longanyang, the official format line uses a Chinese full stop while its example line uses an ASCII period. The request below follows the format line and has not been executed. If your route specifies a more exact requirement, use it in a short test before any batch. Do not treat this example as a verified production recipe.
Separate spoken language from instruction language
The text must use a language the voice supports. The control instruction can have a different requirement: longanyang's documentation requires Chinese instructions even when synthesizing English text.
An unsubmitted APIXO request-body example is:
{
"request_type": "async",
"input": {
"mode": "speech",
"voice": "longanyang",
"prompt": "Welcome. We are glad you are here.",
"instruction": "你说话的情感是happy。",
"format": "mp3"
}
}
Keep the Chinese control string unchanged. The official format uses a Chinese full stop, while its displayed examples use an ASCII period. Follow the voice-specific specification and verify a short sample rather than treating this untested body as a guaranteed recipe.
The APIXO CosyVoice Docs currently limit instruction to 100 characters. Test one emotion before combining role, scenario and language directions.
SSML is a separate capability
A voice can support SSML without supporting Instruct. Confirm each capability separately before combining them.
Use two short sentences as a baseline without an instruction, then repeat the same text and voice with one instruction. An unchanged listening result is not enough to diagnose a country restriction.
What deployment-region limits mean
Alibaba Cloud's region documentation describes access points, deployment and resource configuration. Beijing and Singapore require the appropriate service configuration.
Current official tables list an international Singapore deployment for V3 Flash, and longanyang is listed in both Beijing and Singapore. The reviewed documentation does not support a blanket conclusion that users outside China cannot use the model through an API. This is not a promise that every account, country or network has unrestricted access.
Specific features do differ: the voice-cloning guide states that Singapore V3 Flash does not support synthesis with cloned voices. That feature restriction is different from preset-voice availability. The official deployment table is a supplier reference, not APIXO pricing.
Diagnose APIXO using its own contract
Use APIXO's unified interface and configuration. Its current Docs list speech, clone and design, including preset voices or created voice_id values for synthesis. APIXO's upstream deployment region was not verified, so do not assume a particular Alibaba region or transfer its limits automatically.
Check the voice, text language, instruction length and custom-voice readiness. For failures, retain the task ID, error code and parameters. No cross-country API tests were performed for this guide.
Start with a voice explicitly supporting Instruct, a supported spoken language and one voice-specific instruction. Compare against the baseline, then investigate configuration or routing if necessary. Use APIXO Pricing for platform costs.
API usage rates; Studio monthly credits are separate.
Build with APIXO
Explore image, video, and audio models in one place.
