cstr/bananamind-tts-GGUF

BananaMind-TTS-V2.1 — GGUF

GGUF quantized versions of Banaxi-Tech/BananaMind-TTS-V2.1-Preview for use with CrispASR.

Model overview

BananaMind-TTS-V2.1 is a lightweight Tacotron-lite text-to-speech model (13M params) with a HiFi-GAN vocoder. Character-based tokenizer, 22050 Hz mono output, fixed voice per locale.

Architecture: Embedding → 3× Conv1d+BatchNorm+ReLU → BiLSTM encoder → autoregressive GRU decoder with location-sensitive attention → 5-layer postnet → HiFi-GAN vocoder.

Locales: English (en-us, trained on LJ Speech) and German (de-de, trained on ThorstenVoice).

Files

File	Locale	Quant	Size
bananamind-tts-en-q8_0.gguf	English	Q8_0	~39 MB
bananamind-tts-de-q8_0.gguf	German	Q8_0	~39 MB
bananamind-tts-en-f32.gguf	English	F32	~50 MB
bananamind-tts-de-f32.gguf	German	F32	~50 MB

Q8_0 is recommended for most use cases — negligible quality loss vs F32 at 22% smaller size.

Usage with CrispASR

# English
crispasr -m bananamind-tts-en-q8_0.gguf \
  --tts "This is BananaMind TTS version two point one." \
  --tts-output output.wav

# German
crispasr -m bananamind-tts-de-q8_0.gguf \
  --tts "Hallo, das ist BananaMind TTS Version zwei Punkt eins." \
  --tts-output output_de.wav

Auto-detection works — no --backend flag needed.

Python

import crispasr

session = crispasr.Session("bananamind-tts-en-q8_0.gguf")
pcm = session.synthesize("Hello, world!")
session.save_wav("output.wav", pcm)

Conversion

Converted from the original safetensors using:

python models/convert-bananamind-tts-to-gguf.py \
  --model Banaxi-Tech/BananaMind-TTS-V2.1-Preview \
  --locale en-us --ftype f32 \
  --output bananamind-tts-en-f32.gguf

crispasr-quantize bananamind-tts-en-f32.gguf bananamind-tts-en-q8_0.gguf q8_0

Environment variables

Variable	Description
CRISPASR_BANANAMIND_DEBUG=1	Per-step decoder diagnostics (attention index, stop probability, mel range)
BANANAMIND_TTS_BENCH=1	Per-stage timing (encoder, ar_decoder, postnet, vocoder)

License

Apache 2.0 — same as the original model.

Credits

• Model: Banaxi-Tech — BananaMind-TTS-V2.1-Preview
• English training data: LJ Speech Dataset (Keith Ito & Linda Johnson, 2017)
• German training data: ThorstenVoice Dataset 2022.10 (CC0, DOI: 10.5281/zenodo.7265581)
• GGUF conversion & CrispASR backend: CrispASR