Pocket-TTS - Text-to-Speech Benchmark

Prerequisites

• Ubuntu 20.04 (x86_64)
• Git
• Python 3.10
• Picovoice Console Account

Usage

1. Clone the repository:

git clone https://github.com/Picovoice/text-to-speech-benchmark.git

2. Install the dependencies:

pip3 install -r requirements/requirements-pocket-tts.txt

3. Download the picoLLM model

For each benchmark a picoLLM model is required to generate responses from the LLM. The picoLLM model used in the benchmark is llama-3.2-1b-instruct-385 and can be downloaded from Picovoice Console.

4. Run the benchmark:

Hugging face model download commit hash:

• Repo: kyutai/pocket-tts
  • Commit hash: 427e3d61b276ed69fdd03de0d185fa8a8d97fc5b.
• Repo: kyutai/pocket-tts-without-voice-cloning
  • Commit hash:
    • embeddings_v2: 2578fed2380333b621689eaed6fe144cf69dfeb3.
    • tokenizer.model: d4fdd22ae8c8e1cb3634e150ebeff1dab2d16df3.

For core hour ratio & latency metrics:

python3 benchmark.py \
--picovoice-access-key ${PV_ACCESS_KEY} \
--picollm-model-path ${PICOLLM_MODEL_PATH} \
--engine pocket_tts \

For peak memory metric:

for i in 1 2 3 4 5 6 7 8 9 10 20 40 80; do
    python3 benchmark.py \
    --picovoice-access-key ${PV_ACCESS_KEY} \
    --picollm-model-path ${PICOLLM_MODEL_PATH} \
    --engine pocket_tts \
    --test-memory-size-multiple $i \
done

Replace ${PICOLLM_MODEL_PATH} with the path to the model you downloaded. Replace ${PV_ACCESS_KEY} with your AccessKey obtained from Picovoice Console. Everyone who signs up for Picovoice Console receives a unique AccessKey.