FC-TTS: Diverse Strategies for Speech Synthesis

• FC-TTS is a term that encompasses multiple TTS approaches defined by specialized conditioning strategies, including fully convolutional, flow-consistency, context-aware, and domain-specific adaptations.
• The methods range from non-autoregressive, fully parallel architectures and fine- versus coarse-grained dialogue modeling to flow-matching techniques and dual-reference control to independently manage style and timbre.
• Comparative evaluations reveal trade-offs in latency, quality, and control metrics, emphasizing that FC-TTS must be interpreted in context with each paper’s unique methodology.

FC-TTS designates several distinct strands of text-to-speech research rather than a single canonical architecture. In recent usage, the label has referred to fully convolutional and fully parallel synthesis, fine- and coarse-grained conversational context modeling, flow-consistency formulations for few-step generation, an Arabic football-commentary pipeline with background ambience, and a dual-reference zero-shot framework for independent style and timbre control. What unifies these systems is not a shared backbone, but a shared emphasis on specialized conditioning or generation strategies for speech synthesis (Ma et al., 2018, Hu et al., 2022, Park et al., 20 Jun 2025, Baali et al., 2023, Lee et al., 23 May 2026).

1. Terminological scope

In the literature represented here, FC-TTS is used in multiple senses. The term can denote an architectural family, a context-modeling strategy, a generative-training formulation, or a domain-specific application. This terminological breadth is technically significant because the corresponding systems optimize different objectives, consume different inputs, and report different metrics.

Usage of FC-TTS	Representative system	Central mechanism
Fully convolutional / fully parallel TTS	FPETS (Ma et al., 2018), Fast DCTTS (Kang et al., 2021)	Convolution-only text-to-mel modeling, with UFANS or group-highway designs
Fine- and coarse-grained conversational TTS	FCTalker (Hu et al., 2022)	Joint word-level and utterance-level dialogue context conditioning
Flow-consistency TTS	RapFlow-TTS (Park et al., 20 Jun 2025)	Consistency flow matching with straight-flow and velocity-consistency constraints
Flow-matching TTS without CFG	MG-CFM on F5-TTS (Liang et al., 29 Apr 2025)	Training target reformulated to approximate the CFG trajectory directly
Football-commentator TTS with ambience	FOOCTTS (Baali et al., 2023)	Arabic VITS adaptation plus crowd-noise mixing
Zero-shot style/timbre disentanglement	FC-TTS (Lee et al., 23 May 2026)	Separate style and timbre references with factorized codec conditioning

The most important practical consequence is that “FC-TTS” has to be interpreted locally. In one paper it identifies a convolutional design principle; in another it denotes flow consistency; in another it names a dual-reference controllable zero-shot system. Direct comparison therefore requires attention to the paper-specific definition rather than the acronym alone.

2. Fully convolutional and fully parallel interpretations

The earliest sense is architectural. FPETS defines fully convolutional TTS as a system in which the encoder, alignment mechanism, and acoustic decoder are implemented without recurrent modules, and in which generation is non-autoregressive so that all acoustic frames are predicted simultaneously (Ma et al., 2018). Its encoder is a stack of embeddings, a dense layer, three convolutional layers with kernel size 3 and 1024 channels, and a final dense layer. The alignment module predicts phoneme-wise alignment widths $r_i$ using UFANS, and the decoder is another UFANS network.

A central construction in FPETS is the trainable position encoding. If the input has $N$ phonemes, the alignment module predicts nonnegative widths $r = [r_0,\dots,r_{N-1}]$, from which the absolute positions are defined as

$s_i = \sum_{k=0}^{i-1} r_k + \frac{1}{2}r_i.$

Sinusoidal encodings of phoneme positions and frame positions are then combined into an alignment matrix

$A = F P^\top.$

Because $s_i$ depends on the predicted $r_i$, the positions are learnable through the alignment network rather than fixed at discrete token indices. The resulting alignment is monotonic and near-diagonal, and the paper derives the relation $w_i = \frac14 (r_{i-1} + 2r_i + r_{i+1})$ between attention width and predicted widths (Ma et al., 2018).

FPETS uses a two-stage training strategy. Stage 1 replaces the powerful UFANS decoder with a simpler three-layer CNN decoder so that alignment learning cannot be masked by decoder capacity; Stage 2 freezes the alignment module and reinstates the six-layer UFANS decoder for higher-quality acoustic prediction. On the reported hardware, synthesizing 1 s of audio required 9.9 ms for FPETS, compared with 6157.3 ms for Tacotron2, 494.3 ms for DCTTS, and 105.4 ms for Deep Voice 3. With Griffin–Lim, FPETS obtained MOS $3.65 \pm 0.082$, compared with $3.51 \pm 0.070$ for Tacotron2, $N$0 for DCTTS, and $N$1 for Deep Voice 3. On 100 out-of-domain sentences, FPETS also reported fewer repeats, mispronunciations, and skips than those baselines (Ma et al., 2018).

Fast DCTTS preserves the fully convolutional design principle but not full parallelism. It remains autoregressive over mel frames while reducing cost through narrower and shallower convolutional stacks, group highway activation, pruning, and the EMCD metric for fidelity evaluation (Kang et al., 2021). Its final configuration uses 657,728 parameters and 4,835,728,000 computations, versus 23,896,064 parameters and 275,098,419,200 computations for the baseline DCTTS instantiation. On a single CPU thread, synthesis time decreased from 6.35 s to 0.92 s, a 7.45× speedup. MOS improved from 2.42 to 2.45 on LJSpeech and from 2.62 to 2.74 on KSS. The paper also shows that residual-only replacements destabilize attention, whereas grouped highway gating retains enough gating structure to suppress skipping and repeating more effectively than pure residual connections (Kang et al., 2021).

3. Fine- and coarse-grained conversational context modeling

A different use of FC-TTS appears in FCTalker, where it denotes fine- and coarse-grained context modeling for conversational TTS (Hu et al., 2022). The task is to synthesize an utterance whose linguistic and affective prosody is appropriate to the dialogue context. The paper’s central claim is that utterance-level context alone is insufficient: specific words in prior turns can directly affect how words in the current utterance should be realized.

FCTalker is built on a FastSpeech2-style non-autoregressive backbone and conditions synthesis on four 256-dimensional embeddings: phoneme embeddings $N$2, speaker code $N$3, coarse context $N$4, and fine-grained dialogue context $N$5. The fine-grained encoder is based on a dialogue BERT model with the TOD-BERT base configuration: 12 Transformer layers, 12 attention heads, and hidden size 768. Its input sequence is

$N$6

so that self-attention can model cross-turn token dependencies while the $N$7 markers encode turn identity. A linear layer reduces current-utterance token representations from 768 to 256 to form $N$8. The coarse-grained encoder uses sentence-level BERT embeddings, appends a one-hot speaker ID, aggregates dialogue history with a GRU, and combines the history feature with the current utterance embedding to produce $N$9.

The fusion strategy is deliberately simple: $r = [r_0,\dots,r_{N-1}]$0 is concatenated and passed to the acoustic decoder. Duration, pitch, and energy are handled by the FastSpeech2 variance adaptor, while alignment is obtained through a CTC-based aligner rather than an external forced aligner. The fine-grained dialogue encoder is pretrained with masked language modeling and a dialogue contrastive loss,

$r = [r_0,\dots,r_{N-1}]$1

using approximately 100,000 dialogues across more than 60 domains, and is then fine-tuned within the TTS system (Hu et al., 2022).

Experiments use the DailyTalk conversational corpus, which contains 23,773 audio clips, about 20 hours of speech, and two speakers recorded simultaneously. For $r = [r_0,\dots,r_{N-1}]$2, utterance-level MOS was $r = [r_0,\dots,r_{N-1}]$3 for FastSpeech2, $r = [r_0,\dots,r_{N-1}]$4 for DailyTalk, $r = [r_0,\dots,r_{N-1}]$5 for FCTalker, and $r = [r_0,\dots,r_{N-1}]$6 for ground truth. Dialogue-level MOS was $r = [r_0,\dots,r_{N-1}]$7, $r = [r_0,\dots,r_{N-1}]$8, $r = [r_0,\dots,r_{N-1}]$9, and $s_i = \sum_{k=0}^{i-1} r_k + \frac{1}{2}r_i.$0, respectively. When the history length $s_i = \sum_{k=0}^{i-1} r_k + \frac{1}{2}r_i.$1 was varied from 2 to 14, MOS increased overall from 2 to 12 and then declined slightly from 12 to 14; stability above 4.0 was observed for $s_i = \sum_{k=0}^{i-1} r_k + \frac{1}{2}r_i.$2–12. Reported limitations include degradation for very long histories, domain mismatch between task-oriented dialogue pretraining and open conversational styles such as sarcasm or humor, and the absence of explicit prosody control (Hu et al., 2022).

4. Flow-consistency and guidance-free flow matching

In another line of work, FC-TTS refers to flow-consistency TTS. RapFlow-TTS defines this setting as acoustic modeling that combines flow matching with consistency constraints so that speech can be synthesized with very few ODE integration steps (Park et al., 20 Jun 2025). The underlying ODE is

$s_i = \sum_{k=0}^{i-1} r_k + \frac{1}{2}r_i.$3

and standard flow matching learns a velocity field by regressing $s_i = \sum_{k=0}^{i-1} r_k + \frac{1}{2}r_i.$4 to the ground-truth field. RapFlow-TTS extends this with a consistency flow matching objective

$s_i = \sum_{k=0}^{i-1} r_k + \frac{1}{2}r_i.$5

where $s_i = \sum_{k=0}^{i-1} r_k + \frac{1}{2}r_i.$6 enforces straight-flow consistency and $s_i = \sum_{k=0}^{i-1} r_k + \frac{1}{2}r_i.$7 enforces velocity consistency across neighboring timesteps. The paper further divides time into $s_i = \sum_{k=0}^{i-1} r_k + \frac{1}{2}r_i.$8 segments, trains straight flows in a first stage, adds velocity consistency in a second stage with the encoder frozen, linearly schedules $s_i = \sum_{k=0}^{i-1} r_k + \frac{1}{2}r_i.$9 from 0.1 to 0.001 across $A = F P^\top.$0 bins, and finally adds adversarial learning with a 3:1:2 ratio for $A = F P^\top.$1 (Park et al., 20 Jun 2025).

RapFlow-TTS uses the Matcha-TTS backbone with a text encoder, MAS-based aligner, conditional FM decoder, HiFi-GAN V1 vocoder, and Euler sampling. On LJSpeech with $A = F P^\top.$2, RapFlow-TTS achieved RTF 0.031, WER $A = F P^\top.$3, and MOS $A = F P^\top.$4; RapFlow-TTS† improved these to WER $A = F P^\top.$5 and MOS $A = F P^\top.$6. On VCTK, RapFlow-TTS† at 2 steps obtained WER $A = F P^\top.$7 and MOS $A = F P^\top.$8. The paper characterizes this as a 5-fold reduction in synthesis steps relative to conventional FM approaches and a 10-fold reduction relative to score-based approaches, while noting that many-step synthesis can show limited gains or even WER degradation (Park et al., 20 Jun 2025).

A related but distinct efficiency direction removes inference-time classifier-free guidance from flow-matching TTS (Liang et al., 29 Apr 2025). Instead of enforcing consistency on the trajectory, this work reformulates the conditional FM objective so that the conditional predictor directly approximates the CFG-optimized velocity field. The proposed MG-CFM objective introduces a stop-gradient difference term $A = F P^\top.$9 so that inference no longer requires unconditional evaluation. The model therefore uses two forward passes during training but only a single conditional forward pass per sampling step at inference. Implemented on the F5-TTS base configuration with 335.8M parameters and trained on the 585-hour LibriTTS corpus, the method reported $s_i$0 results of SIM-O 0.597, WER 1.96%, RTF 0.17, and MOS 4.159; at $s_i$1, it reported SIM-O 0.601, WER 2.02%, RTF 0.04, and MOS 4.101, compared with baseline F5-TTS with CFG at $s_i$2: SIM-O 0.592, WER 2.28%, RTF 0.31, and MOS 4.026 (Liang et al., 29 Apr 2025).

5. Domain-specific FC-TTS for Arabic football commentary

FOOCTTS applies the FC-TTS label to a football-commentary pipeline that generates Arabic speech together with the acoustic environment of a live broadcast (Baali et al., 2023). The system is designed for extremely limited in-domain data: approximately 15 minutes of a single Tunisian commentator harvested from YouTube. Its goal is not merely speech synthesis, but synthesis with realistic crowd noise and commentator-like delivery.

The training pipeline extracts audio at 22 kHz, applies voice activity detection with inaSpeechSegmenter, transcribes speech using a large pre-trained Arabic code-switch ASR, transliterates French tokens into Arabic via the QCRI transliteration API, diacritizes transcripts with Farasa, assigns a small set of style labels—neutral, excited, very excited—and then performs CTC segmentation to create aligned training pairs. The TTS model is VITS. It is first pretrained on 1 hour of a male anchor speaker from QASR and then fine-tuned on the commentator data. At inference, raw text is normalized and diacritized, VITS generates clean speech, and the result is mixed with crowd noise under controlled SNR, with optional reverb or EQ (Baali et al., 2023).

Several technical features are specific to Arabic broadcast speech. The frontend is character-based with explicit Arabic diacritics; the paper emphasizes that diacritization is essential because raw Arabic orthography omits short vowels, tanwin, shadda, and many other pronunciation cues. It also notes that grammatical diacritization may diverge from spontaneous spoken realization, so training transcripts are further vowelized “to match speech.” The CTC formulation is used both for alignment and for splitting long commentary into shorter, more homogeneous single-emotion segments (Baali et al., 2023).

The evaluation is qualitative rather than quantitative. The paper reports that, on a new text from a previous game, the generated commentary was “hard to tell that this was generated by a model,” but it does not report MOS, WER/CER, ASR accuracy, or latency. This makes FOOCTTS methodologically different from the benchmark-driven FC-TTS usages in conversational, flow-consistency, or zero-shot controllable TTS.

6. Dual-reference zero-shot control of style and timbre

The most recent use of the term is the 2026 zero-shot system "FC-TTS: Style and Timbre Control in Zero-Shot Text-to-Speech with Disentangled Speech Representations" (Lee et al., 23 May 2026). Here FC-TTS is a dual-reference framework: one reference utterance supplies speaker timbre and another supplies speaking style. The problem is framed as disentangled control under zero-shot conditions, with the explicit objective of keeping style and timbre independently manipulable.

The architecture is organized as a two-stage spectrogram pipeline. A pre-trained FACodec encoder extracts a global speaker embedding $s_i$3 from the timbre reference and prosody tokens $s_i$4 from the style reference, while deliberately excluding FACodec content tokens $s_i$5 and detail tokens $s_i$6 to reduce leakage. Stage 1 is a timbre-anchoring network: a Transformer timbre adapter with AdaLN conditioned on $s_i$7 produces a blurry spectrogram $s_i$8 optimized with an MAE loss. Stage 2 refines $s_i$9 into a clean spectrogram $r_i$0 using a DiT-based flow-matching decoder with 12 DiT blocks and adaLN-Zero, conditioned on hierarchical style embeddings. These style embeddings are built by TCF modules that combine Transformer encoding, Q-Former-style cross-attention, and finite scalar quantization at phoneme and frame levels (Lee et al., 23 May 2026).

The system also introduces a Conditional Consistency Loss. Prosody and timbre predictors operate on generated spectrograms while being cross-conditioned on the non-target attribute, producing

$r_i$1

This loss is intended to reduce ambiguity and leakage between the two control pathways. Training uses LibriHeavy subsets, AdamW with learning rate 0.0002, batch size 64, 200k iterations, 8 V100 GPUs, and about 116 hours. The model has approximately 204M parameters; inference uses 8 NFEs for duration prediction and 32 NFEs with CFG scale 4.0 for spectrogram generation (Lee et al., 23 May 2026).

On LibriSpeech test-clean, FC-TTS reported UTMOS 4.22, WER 1.88, and SPK 0.60, compared with 4.03, 3.30, and 0.67 for a retrained F5-TTS and 4.30, 1.81, and 0.67 for NaturalSpeech 3. In timbre-control tests on RAVDESS, FC-TTS achieved UTMOS 4.03, SPK 0.48, WER 0.18, and ABX timbre preference 66.1%, versus UTMOS 3.19, SPK 0.27, WER 8.40, and ABX 10.7% for FACodec-VC under mismatched conditioning. In prosody-control tests against single-reference F5-TTS, FC-TTS reported UTMOS 3.95, SPK 0.47, WER 0.30, MCD 3.21, ABX 65.5%, and AudioLLM Style-MOS 3.92, whereas F5-TTS reported UTMOS 3.40, SPK 0.57, WER 4.39, MCD 3.43, ABX 8.9%, and Style-MOS 1.50. Ablation results show that removing the entire CCL causes substantial degradation, including LibriSpeech WER 5.88 and RAVDESS WER 9.36 (Lee et al., 23 May 2026).

7. Methodological distinctions and recurrent sources of confusion

A recurrent misconception is that FC-TTS names a single model class. In fact, these works span single-speaker fully convolutional synthesis, dialogue-conditioned FastSpeech2 variants, few-step flow models, Arabic commentator adaptation pipelines, and dual-reference zero-shot control (Ma et al., 2018, Hu et al., 2022, Park et al., 20 Jun 2025, Baali et al., 2023, Lee et al., 23 May 2026). The acronym therefore cannot be interpreted without the surrounding paper.

A second misconception is that FC-TTS necessarily implies either full convolutionality or non-autoregressive generation. FPETS is both fully convolutional and non-autoregressive; Fast DCTTS is fully convolutional but autoregressive; FCTalker uses BERT, GRU, and FFT-based FastSpeech2 components; RapFlow-TTS is defined by consistency-constrained flow matching; the 2026 FC-TTS relies on FACodec, Transformer adapters, Q-Former-style bottlenecks, and DiT blocks (Kang et al., 2021, Hu et al., 2022, Park et al., 20 Jun 2025, Lee et al., 23 May 2026).

A third source of confusion concerns evaluation. Reported outcomes are not directly commensurate across these usages. FPETS emphasizes latency and attention robustness on LJSpeech; FCTalker emphasizes dialogue-level MOS on DailyTalk; RapFlow-TTS emphasizes WER, MOS, NISQA, and RTF on LJSpeech and VCTK; FOOCTTS provides qualitative evidence only; the 2026 FC-TTS emphasizes UTMOS, WER, SPK, MCD, ABX, and AudioLLM judgments on LibriSpeech and RAVDESS (Ma et al., 2018, Hu et al., 2022, Park et al., 20 Jun 2025, Baali et al., 2023, Lee et al., 23 May 2026). This suggests that comparisons should be made within each local definition of FC-TTS rather than across the acronym as such.

Taken together, the literature uses FC-TTS less as a stable taxonomy than as a paper-specific label for a targeted intervention in speech synthesis: full convolutionality and parallelism, contextual prosody modeling, few-step flow generation, domain adaptation with acoustic environment, or disentangled dual-reference control.