faster_tts

Text-to-speech with Drama 3 on CPU and GPU, powered by CTranslate2. Provides a Python library API and the faster-tts command-line interface (CLI). Includes model downloading and caching, sentence splitting, speaker tags, and WAV output. Requires CTranslate2 4.8.4+.

Quick start

Python API:

from faster_tts import Model

model = Model()
result = model.synthesize(
    "Hello! Welcome to our application.",
    output="hello.wav",  # Save the audio as a WAV file.
)

Model() downloads and caches jordimas/drama-3-ct2. Reuse it across calls.

CLI (also available as python -m faster_tts):

faster-tts --text "Hello!" --output hello.wav

Generate emotion, dialogue, and multilingual audio with the sample scripts, or narrate a UTF-8 article:

python samples/article_voice.py article.txt --output article.wav

The article sample selects the device automatically. Use --device cpu or --device cuda to override it. Its default audio budget is 120 seconds (--max-duration-s); reaching it saves partial narration.

Models and devices

model = Model("drama-3")  # Default checkpoint
model = Model("jordimas/drama-3-ct2", revision="main", cache_dir="./models")
model = Model("/path/to/converted-model")
model = Model(local_files_only=True)  # Cached assets only

model = Model(device="cuda", device_index=0, compute_type="float16")
model = Model(device="cpu", compute_type="int8")

The default device="auto" selects GPU when available, otherwise CPU. CUDA and ROCm both use "cuda" with a matching CT2 build. Compute types depend on the build; "default" lets CT2 choose. Inspect settings with model.info(). Set CT2_DRAMA_GPU_WAVEFORM=0 to decode waveforms on CPU.

The CLI accepts the same model and cache options. For local models, use explicit paths such as ./model to avoid interpreting missing directories as Hub IDs. Custom checkpoints must match the engine's layout.

Memory

GPU expert caches size automatically. Set a total budget in decimal GB:

model = Model(
    device="cuda", num_workers=2,
    engine_options={"gpu_expert_cache_gb": 12},
)

This gives each worker a 6 GB cache. Use "auto" or "uncapped" instead of a number, or CLI --gpu-expert-cache-gb. Without an explicit budget, CT2_DRAMA3_GPU_EXPERT_CACHE_GB sets the per-worker limit (0 means uncapped). Automatic sizing reserves at least 2 GB and 30% of free memory and fails if memory cannot be queried or is insufficient. Cache budgets do not cap total VRAM.

When device="auto" selects CPU, large checkpoints can use expert memory mapping on Linux/macOS with compute_type="default" or "float32". Override with CT2_DRAMA3_CPU_MMAP=0 or 1. Explicit device="cpu" leaves it off by default. Keep the checkpoint unchanged while loaded; mapping does not cap RAM usage. Both memory features require a supporting CT2 build.

Synthesis options and results

Pass engine options through engine_options. Experimental directed_spans splits explicit breathing events from speech:

result = model.synthesize(
    "Set the bag down. [breathing] We can continue now.",
    engine_options={"directed_spans": True, "max_frames": 534},
)

Speech spans are capped at 256 frames and breathing events at 22, within the overall budget. Events use empty history and do not condition later speech. Listen to check voice, prosody, and event quality. CLI: --directed-spans.

SynthesisResult includes audio, sample rate, and finish_reason. A "length" result may be truncated; the CLI saves it and exits with status 1. "eos" does not guarantee complete speech. Enable INFO logging for faster_tts.engines.drama3 for device and segment details.

See the API reference, python -m faster_tts --help, and CT2 Drama 3 guide for options and conversion instructions.

Benchmark

Drama 3 only. See benchmark instructions to run your own measurements.

Benchmark results (2026-10-24)

CPU — AMD EPYC 7B13 (62 GiB RAM)

Test text: Hello! Welcome to our application.

Times in the CPU table are means. RTF is compute time divided by generated audio duration; lower is better, and 1 means real-time synthesis. Loading and warmup are excluded.

Implementation	Precision	Time (s)	Audio duration (s)	RTF	Peak RAM (GiB RSS)	Status
HF / PyTorch	bfloat16	300.9	2.8	106.2	44.5	success
CT2	bfloat16	—	—	—	—	unsupported (no matching precision)
HF / PyTorch	float16	351.5	3.2	109.7	44.7	success
CT2	float16	—	—	—	—	unsupported (no matching precision)
HF / PyTorch	int8	—	—	—	—	unsupported (no matching precision)
CT2	int8	18.4	2.4	7.6	44.7	success
HF / PyTorch	int8_float32	—	—	—	—	unsupported (no matching precision)
CT2	int8_float32	17.9	2.4	7.4	44.7	success
HF / PyTorch	float32	155.2	3.0	52.2	51.3	success
CT2	float32	37.7	2.5	15.0	45.5	success

GPU — NVIDIA Tesla T4 (15 GiB VRAM)

CT2 4.8.4, 8 CPU threads, one warmup and three timed runs per successful configuration. Same test text as above. Times are means; peak RAM is process RSS and excludes VRAM. Expert cache: auto (10.874 GB), except the successful float16 run (8 GB).

Implementation	Precision	Time (s)	Audio duration (s)	RTF	Peak RAM (GiB RSS)	Status
CT2	float16	—	—	—	—	CUDA OOM during warmup
CT2	float32	34.5	2.5	13.8	34.5	success
CT2	int8	27.3	4.3	6.3	36.6	success
CT2	int8_float16	26.9	3.0	9.0	35.9	success
CT2	int8_float32	27.2	4.3	6.3	36.6	success
CT2	float16	37.9	3.7	10.3	36.0	success

GPU — NVIDIA L4

CT2 4.8.4, 8 CPU threads, a 12 GB expert cache, one warmup and one timed run per configuration. Same test text as above. Loading and warmup are excluded; peak RAM is process RSS and excludes VRAM. Single runs do not measure timing variability.

Implementation	Precision	Time (s)	Audio duration (s)	RTF	Peak RAM (GiB RSS)	Status
CT2	bfloat16	26.5	2.8	9.5	35.5	success
CT2	float16	27.6	3.1	8.9	35.6	success
CT2	float32	29.4	2.5	11.7	34.6	success
CT2	int8	22.1	4.3	5.1	36.6	success
CT2	int8_bfloat16	21.4	2.8	7.6	35.1	success
CT2	int8_float16	22.8	3.0	7.7	35.9	success
CT2	int8_float32	22.1	4.3	5.1	36.7	success

Contact: Jordi Mas jmas@softcatala.org