# TTS audition — quality first

All new samples generated on orbital M1 Air 8GB, CPU-only.
Live Orpheus unchanged. Native Windows has not been tested with these alternatives.

Listening copies use static gain toward -23 LUFS with -1 dBTP headroom, not compression, denoising or speed changes. Originals remain untouched.
Model-native sample rates retained in original FLOAT WAVs; listening copies are 24 kHz mono PCM16.
Timing excludes model download/load and voice conditioning. Cold first utterances can include warmup. RTF < 1 means faster than real time.

| Voice | Model / reference | Samples | Generation / audio (RTF range) |
|---|---|---:|---:|
| A | Pocket TTS 3.1.0 / alba | 3 | 0.23–0.24 |
| B | Chatterbox Nano / alba | 4 | 0.73–0.81 |
| C | Kokoro 0.9.4 / 82M v1.0 / af_heart | 3 | 0.22–0.23 |
| D | Chatterbox Turbo / alba | 4 | 1.64–1.75 |
| E | Pocket TTS 3.1.0 / azelma | 3 | 0.23–0.24 |
| F | Chatterbox Nano / tara-reference | 4 | 0.71–0.74 |
| G | Kokoro 0.9.4 / 82M v1.0 / af_bella | 3 | 0.21–0.24 |
| H | Chatterbox Turbo / tara-reference | 4 | 1.51–1.68 |

## Interpretation

No subjective quality winner has been declared. Listen for metallic sibilants, breath noise, unnatural pauses, missing/repeated words, emotional range, and voice consistency.
Pocket uses its public preset-only English April 2026 checkpoint, unquantized with the upstream default one-step distilled sampler. Custom voice cloning was not tested: the full checkpoint is gated, and no account terms or credentials were changed.
Chatterbox logged a reference mel/token length alignment warning for both references but produced all requested waveforms. The short synthetic Tara reference is only a first cloning trial, not a claim of optimal cloning fidelity.
The first Pocket run failed because this harness wrapped its multi-threaded decoder in torch.inference_mode(). Replacing that wrapper with torch.no_grad() resolved the cross-thread inference-tensor mutation error. Failed-attempt logs are retained separately and excluded from the successful measurements.
Pocket exposes small streaming chunks. Kokoro yields text chunks; this harness sends a full paragraph per call. Chatterbox's tested API returns a complete waveform, so its first-streaming-audio latency is not measured.
Original audio is hash-checked after transfer. Each normalized copy uses static gain only; peak-limited clips may fall short of the -23 LUFS target.

## Attribution

Alba MacKenna: Casual voice recording, CC BY 4.0, https://huggingface.co/kyutai/tts-voices . Azelma: VCTK p303, CC BY 4.0. Reference-conditioned outputs are synthetic derivatives, not original recordings.

## Reproduction

See audition.py, setup.sh, run.sh, model-revisions.txt and the captured *-environment.txt files. Chatterbox source commit: 5de7a54aa4e5e2baadb0182dde554908b48b85c2. Perth commit: ff1c8ac55a976971245cdd53c18d6131ca00d993.
The isolated environments, weights and remote originals occupy approximately 10 GiB at /Users/akujin/Sources/tts-audition.OApxdw. All rendering and temporary sleep inhibitors finished. Existing user applications remained open; swap usage was about 3.6 GiB near the start and 5.8 GiB after the run, so timings are not empty-machine laboratory measurements. No Windows installations or live voice routing changes were made.