Text-to-Speech (TTS)#

This skill supports Gitee AI TTS plus CosyVoice voice feature extraction workflows. It supports fifteen user-facing model choices for TTS:

• audiofly
• chattts
• cosyvoice2
• cosyvoice3
• cosyvoice-300m
• fish-speech-1.2-sft
• index-tts-1.5
• index-tts-2
• glm-tts
• megatts3
• moss-ttsd-v0.5
• qwen-tts
• spark-tts-0.5b
• step-audio-tts-3b
• vibevoice-large

When the user does not specify a model, ask them to choose one. After the model is chosen, only ask for parameters that are relevant to that model.

Usage#

Use the bundled script to generate speech.

bash

python {baseDir}/scripts/perform_tts.py --model cosyvoice2 --text "你好，我是模力方舟。" --voice alloy --api-key YOUR_API_KEY

For CosyVoice-300M voice feature extraction (voice cloning prep), use:

bash

python {baseDir}/scripts/perform_voice_feature_extraction.py --model FunAudioLLM-CosyVoice-300M --prompt "提供用于声纹提取的提示文本" --file-url "https://example.com/sample.mp3" --api-key YOUR_API_KEY

Options#

• --model required: audiofly, chattts, cosyvoice2, cosyvoice3, cosyvoice-300m, fish-speech-1.2-sft, index-tts-1.5, index-tts-2, glm-tts, megatts3, moss-ttsd-v0.5, qwen-tts, spark-tts-0.5b, step-audio-tts-3b, or vibevoice-large
• --text required in general: text to synthesize. For Qwen3-TTS multi-input mode ( --qwen-inputs-json), --text is optional
• --mode optional: auto, sync, or async
• --prompt optional: model-specific style prompt such as ChatTTS tags
• --prompt-text optional: reference transcript for style-conditioned models
• --prompt-audio-url optional: reference audio URL for style-conditioned models
• --qwen-inputs-json optional: structured Qwen3-TTS inputs JSON (array/object). Supports mixed built-in and custom voice items
• --speaker optional: Qwen3-TTS built-in speaker for single input ( Vivian, Serena, Uncle_Fu, Dylan, Eric, Ryan, Aiden, Ono_Anna, Sohee)
• --language optional: Qwen3-TTS language for single input ( Chinese or English)
• --instruction optional: Qwen3-TTS style instruction for single input
• --prompt-audio-urls optional: vibevoice-large reference audio; supports one URL or JSON array string such as ["https://a.wav","https://b.wav"]
• --emo-audio-prompt-url optional: emotion reference audio URL for IndexTTS-2
• --emo-alpha optional: emotion mixing weight for IndexTTS-2 audio emotion control
• --emo-text optional: emotion control text for IndexTTS-2
• --use-emo-text optional: enable or disable emo_text for IndexTTS-2 ( true/ false)
• --prompt-wav-url optional: reference prompt WAV URL for CosyVoice2 or CosyVoice3
• --voice-url optional: reference voice audio URL for ChatTTS or fish-speech-1.2-sft cloning
• --instruct-text optional: model-specific instruction text such as CosyVoice2 or CosyVoice3 speaking style guidance
• --seed optional: model-specific seed value such as CosyVoice2 or CosyVoice3
• --audio-mode optional: single or role for moss-ttsd-v0.5 (required when mode cannot be inferred from fields)
• --prompt-audio-single-url optional: single-speaker reference audio URL for moss-ttsd-v0.5 single mode
• --prompt-text-single optional: single-speaker reference transcript for moss-ttsd-v0.5 single mode
• --prompt-audio-1-url optional: speaker-1 reference audio URL for moss-ttsd-v0.5 role mode
• --prompt-text-1 optional: speaker-1 reference transcript for moss-ttsd-v0.5 role mode
• --prompt-audio-2-url optional: speaker-2 reference audio URL for moss-ttsd-v0.5 role mode
• --prompt-text-2 optional: speaker-2 reference transcript for moss-ttsd-v0.5 role mode
• --use-normalize optional: enable or disable use_normalize for moss-ttsd-v0.5 ( true/ false)
• --prompt-language optional: prompt language hint for models such as MegaTTS3
• --intelligibility-weight optional: pronunciation intelligibility weight for models such as MegaTTS3
• --similarity-weight optional: timbre similarity weight for models such as MegaTTS3
• --temperature optional: model-specific sampling temperature
• --top-p optional: model-specific top-p sampling value
• --top-k optional: model-specific top-k sampling value
• --gender optional: async TTS gender hint
• --pitch optional: async TTS pitch hint
• --speed optional: async TTS speed hint (for example CosyVoice3, Spark-TTS-0.5B, or Qwen3-TTS)
• --num-inference-steps optional: AudioFly generation step count
• --guidance-scale optional: AudioFly classifier-free guidance scale
• --output-format optional: AudioFly or Qwen3-TTS output format such as mp3 or wav
• --voice optional: OpenAI-compatible voice field when supported by the target model
• --extra-body-json optional: JSON object for explicitly requested undocumented fields
• --response-data-format optional: url or blob for sync TTS
• --output optional: output file path when sync TTS returns binary audio
• --failover-enabled optional: request header X-Failover-Enabled, defaults to true
• perform_voice_feature_extraction.py options: --prompt, --file-url (URL only), --model (default FunAudioLLM-CosyVoice-300M), --failover-enabled, --output, --api-key

Workflow#

1. Determine whether the user wants speech synthesis or CosyVoice voice-feature extraction.
2. For speech synthesis: ask the user to choose one of audiofly, chattts, cosyvoice2, cosyvoice3, cosyvoice-300m, fish-speech-1.2-sft, index-tts-1.5, index-tts-2, glm-tts, megatts3, moss-ttsd-v0.5, qwen-tts, spark-tts-0.5b, step-audio-tts-3b, or vibevoice-large if not specified.
3. For speech synthesis: read references/models.md, gather missing model-specific params, and execute perform_tts.py.
4. For voice-feature extraction: execute perform_voice_feature_extraction.py with --prompt and URL-only --file-url.
5. Parse script output.
6. For TTS output, prioritize AUDIO_URL: then AUDIO_FILE: then TTS_RESULT:.
7. For voice feature output, prioritize VOICE_URL: (if present), otherwise return VOICE_FEATURE_FILE: and summarize VOICE_FEATURE_RESULT:.

Notes#

• Keep the answer language consistent with the user's language.

• This script is standard-library only and is intended to run directly with python; do not require uv for moark-tts.

• If GITEEAI_API_KEY is missing, remind the user to provide --api-key.

• By default, all TTS requests send X-Failover-Enabled: true. Only set --failover-enabled false when the user explicitly needs to disable failover.

• audiofly is mapped to the official model name AudioFly. Use async mode only. When the user shows an OpenAI SDK example that puts num_inference_steps, guidance_scale, or output_format under extra_body, map them to --num-inference-steps, --guidance-scale, and --output-format.

• chattts is mapped to the official model name ChatTTS. When the user shows an OpenAI SDK example that puts prompt, temperature, top_P, top_K, or voice_url under extra_body, map them to --prompt, --temperature, --top-p, --top-k, and --voice-url.

• cosyvoice2 is mapped to the official model name CosyVoice2. When the user shows an OpenAI SDK example that puts prompt_wav_url, prompt_text, instruct_text, or seed under extra_body, map them to --prompt-wav-url, --prompt-text, --instruct-text, and --seed.

• cosyvoice3 is mapped to the official model name CosyVoice3. Use async mode only. When the user shows an OpenAI SDK example that puts prompt_wav_url, prompt_text, instruct_text, speed, or seed under extra_body, map them to --prompt-wav-url, --prompt-text, --instruct-text, --speed, and --seed.

• cosyvoice-300m is mapped to FunAudioLLM-CosyVoice-300M for sync /audio/speech. Map OpenAI extra_body.voice_url to --voice-url.

• CosyVoice voice-feature extraction uses /audio/voice-feature-extraction and is handled by perform_voice_feature_extraction.py; --file-url must be an http(s) URL (no local file path support).

• fish-speech-1.2-sft uses sync /audio/speech. When the user shows an OpenAI SDK example that puts voice_url under extra_body, map it to --voice-url.

• index-tts-1.5 currently uses the sync /audio/speech endpoint. When the user shows an OpenAI SDK example that puts prompt_audio_url under extra_body, map it to the script's --prompt-audio-url.

• index-tts-2 supports four emotion-control patterns: sync/async + audio-emotion/text-emotion. Map emo_audio_prompt_url + emo_alpha to --emo-audio-prompt-url + --emo-alpha; map emo_text + use_emo_text to --emo-text + --use-emo-text. In auto mode it defaults to sync; when user asks async, force --mode async.

• megatts3 is mapped to the official model name MegaTTS3. When the user shows an OpenAI SDK example that puts prompt_language, intelligibility_weight, or similarity_weight under extra_body, map them to --prompt-language, --intelligibility-weight, and --similarity-weight.

• step-audio-tts-3b is mapped to the official model name Step-Audio-TTS-3B. When the user shows an OpenAI SDK example that puts prompt_audio_url and prompt_text under extra_body, map them to --prompt-audio-url and --prompt-text.

• spark-tts-0.5b is mapped to the official model name Spark-TTS-0.5B. Use async mode only. For plain synthesis, just pass text. For voice cloning, map prompt_audio_url and prompt_text to --prompt-audio-url and --prompt-text; gender/ pitch/ speed can be passed when explicitly requested.

• qwen-tts is mapped to the official model name Qwen3-TTS. Use async mode only. Prefer structured inputs items:

  • Built-in speaker item: prompt + speaker + optional language ( Chinese/ English) + optional instruction.
  • Custom voice item: prompt + prompt_audio_url + prompt_text + optional language + optional instruction.
  • Use --qwen-inputs-json for multiple items in one request; use --speaker/ --language/ --instruction for single-item mode.
  • Built-in speakers: Vivian, Serena, Uncle_Fu, Dylan, Eric, Ryan, Aiden, Ono_Anna, Sohee.

• moss-ttsd-v0.5 is mapped to the official model name MOSS-TTSD-v0.5. Use async mode only. Map single mode fields prompt_audio_single_url + prompt_text_single to --prompt-audio-single-url + --prompt-text-single, and role mode fields ( prompt_audio_1_url, prompt_text_1, prompt_audio_2_url, prompt_text_2) to the matching CLI options. Pass audio_mode through --audio-mode and use_normalize through --use-normalize.

• vibevoice-large is mapped to the official model name VibeVoice-Large. Use async mode only. Map prompt_audio_urls to --prompt-audio-urls, and accept both a single URL string and a JSON array string. When the user provides only prompt_audio_url, map it into prompt_audio_urls automatically for compatibility.

• glm-tts currently exposes only the basic sync request in the official OpenAPI spec.

• Do not invent model parameters. If a field is not documented for that model, only pass it when the user explicitly asked for it and use --extra-body-json.